METHOD AND APPARATUS FOR CLASS INCREMENTAL LEARNING

Abstract

The present application generally relates to a method for training a machine learning, ML, model using class incremental learning, and to a computer-implemented method and apparatus for using the trained machine learning, ML, model. The method may learn how to update semantic representations of old concepts (classes) by modelling drift of semantic representations. The method may also learn how to update feature representations of old concepts (classes) by modelling drift of feature representations

Description

CROSS-REFERENCE TO RELATED APPLICATION(S)

This application is based on and claims priority to Great Britain Patent Application No. GB2116074.2, filed on Nov. 9, 2021, and Great Britain Patent Application No. GB2205999.2, filed on Apr. 25, 2022 in the Great Britain Intellectual Property, the disclosures of which are incorporated by reference herein in their entireties.

BACKGROUND

1. Field

The present application generally relates to a method for training a machine learning, ML, model using class incremental learning, and to a computer-implemented method and apparatus for using the trained machine learning, ML, model.

2. Description of the Related Art

Continual learning (also called lifelong learning) refers to the ability to continuously learn and adapt to new environments, exploiting knowledge gained from the past when solving novel tasks. Though being a common human trait, the lifelong learning paradigm hardly applies to artificial intelligence systems. Nevertheless, a classification model can be progressively trained on a constantly changing set of new classes, and this may be termed Class-Incremental Learning (CIL). Continual learning has been extensively studied in a class incremental fashion [7, 22, 23]. CIL may be used in several technical fields, for example image classification.

FIG. 1A is a schematic illustration of a typical image classifier which receives an input image and which may be trained using CIL. Merely as an example, the image classifier may use a general convolution neural network (CNN) architecture, for example as described in “Deep Learning” by LeCun et al. published in Nature Vol. 521, May 2015. As shown in FIG. 1A, a set of features are extracted from within the image using a feature extractor 100. The features may be extracted using any suitable technique, including convolution kernels. The use of convolution kernels is described for example in “Development of convolutional neural network and its application in image classification: a survey” by Wang et al. published in Optical Engineering 2019. As set out above, the feature extractor may have a CNN architecture and may thus comprise a plurality of blocks 102 each of which comprises a plurality of convolutional layers and a reLU layer. There is pooling from the first block 102 to the second block 102 and so on until the final block. In a final stage, there is flattening to generate an output feature representation 104. The feature representation is then passed to a classifier 106 which also comprises a plurality of layers. The output is a probability that the extracted kernel(s) matches one of the classes which is known by the model. In this example, the input image is a zebra and the output is a probability of 0.2 that the image is a horse, 0.7 that the image is a zebra and 0.1 that the image is a dog. Assuming that the threshold for a correct classification is below 0.7, the classifier has correctly classified the image as a zebra.

When training a model such as the one shown in FIG. 1A with CIL, the model may be initially trained with images classified as dogs and at the next time step images classified as zebras or horses may be added. CIL is schematically illustrated in FIG. 1B, and initially at time step t−1, several items or images which are schematically indicated by triangles have the same classification, e.g. an image of a dog. The images form a cluster 112 which is centred on a class prototype 114. Such a class prototype 114 may be considered to be a template of features which represents an average dog. A decision boundary 110 of the feature space within which all images would be classified as a dog is indicated by a solid line. The decision boundary is extracted using the classifier of FIG. 1A.

In the next time step t, several new classified images (i.e. several images belonging to new classes) are introduced and are indicated by squares and circles respectively. For example, the squares may be images of zebras and the circles images of horses. As indicated, some of the horse and zebra images are within the decision boundary 110 of the feature space for dogs and potentially would be mis-classified as dogs. The original classified images are shown as ghost triangles because they are not included when training the model at time step t. In other words, for each incremental training task and step t, the training set is composed of images belonging to the current class sets (e.g. horses and zebras) C_t, whereas past semantic categories (e.g. dogs) C_old ≙{C_t′}_t′=1^t−1 lack any training sample. The final objective for the model is to maximize the generalization (classification) accuracy on all the categories observed up to the end of the stream.

As shown in the final section of FIG. 1B, the training of the model with the new images leads to an adjustment to the original decision boundary 110, and new decision boundaries 120,130 for the new zebra and horse classifications respectively are learnt. New images of dogs, horses and zebras can then be classified by the newly trained classifier. As shown, the new images of horses and zebras indicated by the solid squares and circles are correctly located within the relevant portion of the feature space. However, the lack of images classified as dogs means that the feature representation for the class prototype for dogs has drifted towards the decision boundary between the dog and horse feature spaces. Accordingly, some of the original images which were centred around the class prototype would now be incorrectly classified as horses. These images are circled by the group 140. In other words, the change of distribution of training data D_t in the form of semantic drift (i.e, change of the experienced class set C_t) naturally leads to catastrophic forgetting. It will be appreciated that FIG. 1B illustrates catastrophic forgetting as occurring after a single step but it is likely that the bias towards new data which causes past learned information will be more gradually erased.

The catastrophic forgetting phenomenon has also been described as the stability-plasticity dilemma, since there exists a trade-off between the search for preservation of old-task knowledge (i.e., stability) and the necessity to accommodate for the information obtained from the experience of new tasks (i.e., plasticity). To preserve past knowledge, many state-of-the-art methods store and rehearse training exemplars of previously seen classes, which might be source of privacy or storage issues. Alternatively, some recent approaches resort to prototypical feature representations to inject information of former classes in later incremental steps. However, if not updated, those representations progressively become staler and more outdated as the incremental learning progresses and new classes are learned.

Most of the successful known CIL methods use exemplars of old classes C_old to rehearse past knowledge [1, 4, 5, 8, 9, 12, 18, 20, 25, 27, 32, 36]. However, storing samples belonging to all classes might be impractical due to limited resource availability or privacy requirements. To address CIL without storing exemplars, regularisation methods have been proposed [2, 6, 16, 39, 41]; the common goal is to identify key model parameters to solve old tasks, and prevent their change when learning a new task. Alternatively, knowledge distillation has been proposed [11, 19], where representations of new classes are forced to only slightly deviate from their original version computed at the beginning of the incremental step for learning the current task. Yet, those methods usually underperform when compared to the state-of-the-art (SotA) solutions.

In a more recent work [43], class prototypes were used to inject past knowledge. Although showing promising results, this method fails to capture the representation drift that is present while incrementally training the model and that is illustrated in FIG. 1B. That is because feature representations (e.g. texture, shape and colour and their composition representing parts of objects, scenes etc.) of old classes (e.g. different animals) are getting constantly staler and more outdated as the learning progresses and the model is updated. The estimated feature representation may be represented by F_old{F_t′}_t′=1^t and the old classes by C_old. The old classes may have been computed when the corresponding data is available and then kept fixed for the rest of the training.

A different work [40] proposes to estimate the change of prototypes (e.g. templates of average dogs) of old classes (e.g. dogs) when learning new classes (e.g. horses and zebra). The old templates may be represented by Π_old{_t′}_t′=1^t and the old classes by C_old and the new classes by C_t. However, this method is limited in scope to embedding learning and devises a deterministic non-learnable module to estimate prototype shift.

There are also examples from the patent literature. For example, US2020151619 describes a system and method for accounting for the impact of concept drift in selecting machine learning training methods to address the identified impact. US2021224696 describes a computer system for detecting a concept drift of a machine learning model in a production environment for adaptive optimization of the concept drift detection. US2021157704 describes a system which can monitor applications and analyze the metrics to determine if one or more of the applications are regressing or performing as expected. US2021073627 describes a method for detecting degradation of the machine learning model based on key performance indicators.

The present applicant has recognised the need for an improved system and method that addresses these drawbacks.

SUMMARY

In a first approach of the present techniques, there is provided a computer-implemented method for training a classifier using class incremental learning, CIL, wherein the classifier is a machine learning, ML, model comprising a feature extraction model and a classification model and wherein at each incremental time step in the CIL a new training dataset having a set of samples with class labels is obtained and an optimisation over a plurality of stages is applied. The training method may comprise for each incremental step: extracting, using the feature extraction model, a first set of features representing the samples within the new training set at the beginning of the incremental step and a second set of features representing the samples within the training set at a subsequent stage within the optimisation; learning a feature drift model using the extracted first and second sets of features, wherein the feature drift model estimates the drift between the features representing each sample as the classification model is updated; obtaining a first set of old class prototypes for the beginning of the incremental step, wherein each class prototype is a semantic representation of the sets of features of the samples which have the same class label and wherein each old class prototype represents old samples which are samples that are not present in the new training set; learning a semantic drift model using the set of old class prototypes and first set of features, wherein the semantic drift model estimates the drift between the class prototypes for each class as the classification model is updated; inferring at least one of a second set of old class prototypes and a set of old features using the feature drift model and the semantic drift model, wherein the second set of prototypes are semantic representations of the old samples at a subsequent stage within the optimisation and the set of old features are the feature representations of revived old samples at the subsequent stage within the optimisation; and using at least one of the inferred second set of old class prototypes and set of old features to update the ML model. When the final incremental step is completed, the method may further comprise outputting the trained classification model.

In other words, there is provided a computer-implemented method for training a machine learning, ML, model. More specifically, there may be provided a computer-implemented method for performing class incremental learning of at least one or both of feature and semantic representation using a machine learning, ML, model. More specifically, there may be provided a computer-implemented method for performing class incremental learning using a machine learning, ML, model, the method comprising: inferring a set of prototypes of old classes using at least one of a feature drift model which is used to represent feature drift and a semantic drift model which is used to represent semantic drift; and using the inferred set of prototypes to update the ML model.

The method may learn how to update semantic representations (also termed class prototypes or templates) of old concepts (classes) by modelling drift of semantic representations. The method may include modelling the relationship with the new classes currently being learned for which data samples are available. The method may also learn how to update feature representations of old concepts (classes) by modelling drift of feature representations. The method may include modelling the trajectory of samples in the feature space spanned by parameters of the task network. These two models may be used separately or in combination to re-estimate continually evolving feature distributions of old concepts.

By implementing the modelling described above, this avoids the need to store samples of old classes to access evolving data distributions of old classes. The modelling also avoids using fixed representations of old classes which typically become staler at each new incremental phase. In other words, the techniques described provide a ML framework capable of modelling inter-class relationships and estimating the evolution of semantic and feature representations. The techniques described also provide a framework to achieve high performance on class-incremental learning, without requiring data samples to be stored.

When compared to prior art techniques, e.g. US2020151619, it is noted that our proposed method focuses on representation drift instead of concept drift, that is, we consider change of learned representations of ML models instead of only change of data. Our method may thus be considered as aiming to model change of representation drifts instead of determining their impact on model performance. There is an aim of improving the new representations together with old representations by updating a model, instead of selecting a model from multiple models. For example, considering US2021224696 as a contrast, our proposed method may be described as using a single ML model which is continually updated using novel data and is never retrained on past data. Thus, there is no need to store and/or access past data. For example, considering US2021157704, our method by contrast aims to directly address a continual learning problem where the training data is continually changing and data from the past cannot be accessed. We assume that the training data is accessible for a limited period of time and cannot be stored and reused. When compared to US2021073627, our method aims at directly addressing the performance degradation of an ML model caused by catastrophic forgetting by modelling semantic and/or feature representation drifts. Only new data which is currently available is leveraged.

Each new training dataset may be represented by _t and thus the complete dataset for the CIL may be represented as {_t}_t=0^T=(_t,_t), t=0, . . . , T, where _t={x_t,j}_j=1^Nt is the set of samples, _t={y_t,j∈_t}_j=1^Nt is the set of their labels. _t is the set of class labels observed at this step and N_t is the number of samples per step. Each sample may be an image, text data or audio data. When images are used as data and the ML task is image classification, the ML model may be an image classifier.

The feature drift model may be represented by Γ_γtn at each optimisation stage n and each incremental step t and may be parameterized by a set of parameters γ of the feature drift model. The feature drift model estimates the drift between the features representing each sample in each class as the classification model is updated (i.e. at each incremental step and at each optimisation stage within each step). In other words, the feature drift model may learn the drift experienced by the features of the set of class labels _t observed at each incremental step. The second semantic drift model may be represented by Ψ_ψtn at each optimisation stage n and each incremental step t and may be parameterized by a semantic drift set of parameters ψ. The semantic drift model estimates the drift between the class prototypes for each class as the classification model is updated (i.e. at each incremental step and at each optimisation stage within each step). In other words, the semantic drift model may learn the drift experienced by while moving from the set of old class labels _old to the set of class labels _t observed at each incremental step.

The first set of features representing each sample within the new training set at the beginning of the incremental step may be represented by _t⁰. The second set of features representing each sample within the training set at a subsequent stage within the optimisation (i.e. the nth optimisation stage) may be represented by _tⁿ. The first set of old class prototypes for the beginning of the incremental step may be represented by Π_old^t,0. The second set of old class prototypes for the subsequent stage within the optimisation may be represented by Π_old^t,n. From the old class prototypes at the subsequent stage, we can estimate the set of features for the old samples _oldⁿ which are not part of the training dataset _t. These old samples which are estimated may be termed revived old samples or revived evanescent representations because the representations were no longer appearing in the dataset.

The extracting, learning, obtaining, inferring and using steps may be repeated for each optimisation stage within each incremental step. When learning the feature drift and semantic drift models, the feature extraction and classification models are kept fixed. These learning steps may be termed a learning phase. When using the inferred information to update the ML model (i.e. to update the feature extraction and classification models), the feature drift and semantic drift models are kept fixed. This updating step and preceding inferring step may be termed an inference phase. At each learning and using step, the models may be trained until convergence where the convergence criterion for a model is early stopping the optimisation of model parameters if the training loss does not change for a fixed/set number of steps. There may be an initialisation stage for the first optimisation stage n=0. In the initialisation stage, both the feature extraction model f_θ0 and the classification model h_ϕ0 for the incremental time step 0 may be trained using a loss function L_cc⁰, on the initial dataset ₀. The set of class prototype(s) Π_new⁰ can be computed and then we initialise the first set of old class prototypes Π_old^t=1,0 as Π_new⁰.

Inferring a set of prototypes of old classes using the feature drift model may comprise training the feature drift model using the available set of features; and using the trained feature drift model to estimate the feature drift on past data whereby the set of prototypes of old classes is inferred. Training may be done using any suitable technique and depends on the architecture of the machine learning model.

The feature drift model may use a linear deep neural network (DNN), e.g. a multilayer perceptron (MLP). The feature drift model may be learned by using a feature drift loss function which minimises the error between the extracted second set of features and a third set of features which are the estimate of these features from the first set of features using the feature drift model. Merely as an example, the loss function L_f^t may be expressed using the mean squared error function:

L _f ^t(F_t⁰,F_tⁿ;γ_tⁿ)=L_mse,f^t≙∥Γ_γtn(F_t⁰)−F_tⁿ∥₂²

where F_t⁰ is the extracted first set of features, F_tⁿ is the extracted second set of features and Γ_γtn is the feature drift model.

The feature drift model may use a variational model such as a variational auto-encoder (VAE). Training the feature drift model may comprise updating the weights γ of the model to maximize the likelihood p(F∈_tⁿ|F∈_t⁰; γ) where _t⁰ denotes the set of features extracted at the optimization step 0 (i.e. the extracted first set of features), _tⁿ denotes the set of features extracted using the feature extractor f_Θt updated with n optimization epochs at the incremental step t (i.e. the extracted second set of features) and F is the up-to-date feature representations from both the set of features of new classes and old classes. Using the expression, p(F∈_tⁿ|F∈_t⁰; γ_tⁿ) we can estimate drifts more accurately at each incremental step and the optimisation stage. In this arrangement, the loss function L_f^t for feature drift may be expressed as:

L _f ^t(F_t^o,F_tⁿ;γ_tⁿ)=βL_rec,f^t+(1−α)L_kl,f^t+(α−λ_info−1)L_info,f^t

where L_rec,f^t is the reconstruction loss, L_kl,f^t is the KL divergence loss and L_info,f^t is the loss of the InfoVAE and α, β and λ_info are constants.

Inferring a set of prototypes of old classes using the semantic drift model may comprise training the semantic drift model using the available set of features; and using the trained semantic drift model to estimate the semantic drift on past data whereby the set of prototypes of old classes is inferred. Training may be done using any suitable technique and depends on the architecture of the machine learning model.

The semantic drift model may use a linear deep neural network (DNN), e.g. a multilayer perceptron (MLP). The semantic drift model may be learned by using a semantic drift loss function which minimises the error between the first set of old class prototypes and a third set of old class prototypes which are the estimate of these old class prototypes from the extracted first set of features using the semantic drift model. Merely as an example, the loss function L_s^t for semantic drift may be expressed as:

L _s ^t(F_t^o,Π_old^t,o;ψ_tⁿ)=L_mse,s^t∥Ψ_ψtn(F_t^o)−Π_old^t,o∥₂²

where F_t^o is the extracted first set of features, Π_old^t,o is the obtained first set of old class prototypes and Ψ_ψtn is the semantic drift model.

The semantic drift model may use a variational model such as a variational auto-encoder (VAE). Training the semantic drift model may comprise updating the weights ψ of the model to maximize the likelihood p(π∈Π_old^t,n=0|F∈_t⁰; ψ) where _t⁰ denotes the set of features extracted at the optimization step 0 (i.e. the extracted first set of features), Π_old^t,0 is the set of prototypes of old classes at the optimization step 0 (i.e. the first set of old class prototypes), F is the up-to-date feature representation and π is the up-to-date semantic representation. Using the expression, p(π∈Π_old^t,n=0|F∈_t⁰; ψ_tⁿ) we can more accurately capture the drift at incremental steps and optimisation stages. In this arrangement, the loss function L_s^t for semantic drift may be expressed as:

L _s(F_t^o,Π_old^t,o;ψ_tⁿ)=L_rec,s+λ_kldL_kld,s

where L_rec,s^t is the reconstruction loss, L_kl,s^t is the KL divergence loss and λ_kld,s is a constant set to 1.

The feature drift and semantic drift model may be jointly trained, for example by fusing them to estimate the distribution of revived evanescent representations p(F∈F_oldⁿ) using either architecture. For this purpose, we optimise model parameters by minimising a fusion loss which measures discrepancy between the estimated distributions from either model p(F∈F_oldⁿ; γ_tⁿ) and p(F∈F_oldⁿ; ψ_tⁿ) employing a suitable training objective also termed a fusion loss may be expressed as

L _fus ^t=∥Π_old,s^t,n−Π_old,f^t,n∥₂²+λ_corr∥ρ(Π_old,s^t,n)−ρΠ_old,f^t,n)∥₂²

where the subscript s and f denotes the updated prototypes of old classes estimated by the semantic and feature drift models, respectively, ∥·∥₂² is the squared ₂ norm, λ_corr>0 is the regularisation parameter and ρ(Π) is the normalised correlation matrix of Π[3].

Inferring at least one of a second set of old class prototypes and a set of old features using the feature drift model and the semantic drift model may comprise using the feature drift model and the semantic drift model jointly or individually. The feature drift model Γ_γtn may be used to infer feature drift representations of old class prototypes Π_old,f^t,n at stage n>0 (i.e. feature drift versions of the second set of old class prototypes) using representations of old class prototypes at the beginning of the time step Π_old^t,0 (i.e. the first set of old class prototypes) as the input. The subscript f denotes that the class prototypes are inferred from the feature drift model. These second set of old class prototypes may be used to revive the evanescent representations, i.e. to generate the set of old class features F_oldⁿ at the subsequent stage n of the optimisation. Alternatively, the set of old class features F_oldⁿ at the subsequent stage n of the optimisation may be inferred direct from the feature drift model and/or the semantic drift model. The nature of the inferring may depend on the architecture of the machine learning model. In either architecture, the feature drift model is kept fixed when inferring.

As set out above, the feature drift model may use a linear deep neural network (DNN). In this example, the feature drift model enables directly tracking of the trajectory of the old class prototypes Π_old^t,n as the optimisation proceeds. Thus, the second set of old class prototypes can be mapped from the first set of old class prototypes Π_old^t,0 at the start of the incremental step. The distribution of the features of the revived evanescent representations may then be approximated by a Gaussian distribution from the second set of old class prototypes.

As set out above, the feature drift model may use a variational model and inferring the second set of old class prototypes may comprise using the trained feature drift model to compute the posterior probability p(π∈Π_old^t,n>0|π∈Π_old^t,n=0; γ) where Π_old^t,0 is the first set of prototypes of old classes, Π_old^t,n is the second set of prototypes of old classes at the step t, and γ are the weights of the feature drift model. In other words, the trained feature drift model may be used to approximate the probability that the set of features is a member of the set of features for the revived evanescent representations at the nth optimisation stage given the set of features is a member of the set of features for the revived evanescent representations at the start of the time step. p(F∈_oldⁿ|F∈_old⁰). This distribution of the features of the revived evanescent representations may also be approximated by a Gaussian distribution from the second set of old class prototypes.

The semantic drift model Ψ_ψtn may be used to infer semantic drift representations of old class prototypes Π_old,s^t,n at stage n>0 (i.e. semantic drift versions of the second set of old class prototypes) using representations of old class prototypes at the beginning of the time step Π_old^t,0 (i.e. the first set of old class prototypes) as the input. The subscript s denotes that the class prototypes are inferred from the semantic drift model. The nature of the inferring may depend on the architecture of the machine learning model. In either architecture, the semantic drift model is kept fixed when inferring.

As set out above, the semantic drift model may use a linear deep neural network (DNN). In this example, the semantic drift model enables directly tracking of the trajectory of the old class prototypes Π_old^t,n as the optimisation proceeds. Thus, the second set of old class prototypes can be mapped from the extracted second set of features _tⁿ. The distribution of the features of the revived evanescent representations may then be approximated by a Gaussian distribution from the second set of old class prototypes.

As set out above, the semantic drift model may use a variational model and inferring the second set of old class prototypes may comprise using the trained feature drift model to compute the posterior probability p(π∈Π_old^t,n>0|F∈_tⁿ; ψ), where Π_old^t,n is the set of prototypes updated at the n^th epoch at the step t, _tⁿ denotes the set of features extracted using the feature extractor f_Θt updated with n optimization epochs at the incremental step t and ψ are the weights of the semantic drift model. In other words, the trained semantic drift model may be used to approximate the probability that the set of features is a member of the set of features for the revived evanescent representations at the nth optimisation stage given the set of features is a member of the set of features for the set of features at the subsequent stage of the optimisation process. p(F∈_oldⁿ|F∈_tⁿ). This distribution of the features of the revived evanescent representations may also be approximated by a Gaussian distribution from the second set of old class prototypes.

In other words, a linear model may be used for at least one of the feature drift model and the semantic drift model. Similarly, a non-linear model may be used for at least one of the feature drift model and the semantic drift model.

Using the inferred second set of old class prototypes to update the ML model may comprise training the feature extraction model and the classification model based on at least one of the inferred second set of old class prototypes and the inferred set of old features. The inferred second set of old class prototypes and the inferred set of old features may be used to generate the revived evanescent representation, i.e. the revived old samples. The training dataset for the updating step may include the data in the new dataset so that the feature extraction model and the classification model are trained on the new dataset with the revived evanescent representations. The training may be done by any suitable means but when training the feature extraction model and the classification model, the feature drift and semantic drift models are kept fixed.

The feature extraction model may be trained using a distillation loss which approximates the difference between the overall classification objective L_cc^t and the loss L_pc^t of the previous model f_θt-1 (i.e. the model at the previous incremental time step) on the new dataset _t. The distillation loss may be expressed as the ₂ distance between representations (i.e. set of features) extracted from _t using the current feature extraction model f_θt and the feature extraction model from the previous step f_θt-1.

The classification model may be trained using a cross-entropy loss. The classification model may also be trained using a representation drift loss, for example a representation drift loss function L_rd^t may be calculated using

$L_{\mathrm{rd}}^t\stackrel{\Delta }{=}{L}_{\mathrm{rd}}\left({ℱ}_{\mathrm{old}}^{n,t}\right)=\sum _{F\in {ℱ}_{\mathrm{old}}^{n,t}}{y}_F\log h_{\varphi t}\left(F\right)$

where y_F is the one-hot label vector of the class c_F∈_old and _old^n,t is the set of revived evanescent representations sampled from the estimated distribution using the updated prototypes of _old.

The feature extraction and classification models may be trained together. For example, the overall classification objective L_cc^t computed at each step t may be calculated from:

L _cc ^t(D_t,F_oldⁿ;θ_t,ϕ_t)=L_ce^t+λ_rdL_rd^t+λ_fkdL_fkd^t,

where L_ce^t is the cross-entropy loss, L_rd^t and L_fkd^t are the representation drift and distillation loss are used only for t>0 with loss balancing parameters λ_rd>0 and λ_fkd>0.

In another approach of the present techniques, there is provided a method of classifying a sample, the method comprising inputting a sample, e.g. an image; extracting a plurality of features from the sample using a feature extraction model of a classifier which has been trained according to the method described above; classifying the sample based on the extracted plurality of features using a classification model of a classifier which has been trained according to the method described above; and outputting the result of the classifying step to give a classification of the sample.

In another approach of the present techniques, there is provided an apparatus for performing or modelling drift of learned representations for class-incremental learning using a machine learning, ML, model. The apparatus may further comprise at least one image capture device for capturing images or videos to be processed or classified by the trained ML model. The apparatus may further comprise at least one interface for providing a result of the processing by the ML model to a user of the apparatus.

The apparatus may be any one of: a smartphone, tablet, laptop, computer or computing device, virtual assistant device, a vehicle, a drone, an autonomous vehicle, a robot or robotic device, a robotic assistant, image capture system or device, an augmented reality system or device, a virtual reality system or device, a gaming system, an Internet of Things device, or a smart consumer device (such as a smart fridge). It will be understood that this is a non-exhaustive and non-limiting list of example apparatus.

In a related approach of the present techniques, there is provided a non-transitory data carrier carrying processor control code to implement the methods described herein.

As will be appreciated by one skilled in the art, the present techniques may be embodied as a system, method or computer program product. Accordingly, present techniques may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects.

Furthermore, the present techniques may take the form of a computer program product embodied in a computer readable medium having computer readable program code embodied thereon. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. A computer readable medium may be, for example, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing.

Computer program code for carrying out operations of the present techniques may be written in any combination of one or more programming languages, including object oriented programming languages and conventional procedural programming languages. Code components may be embodied as procedures, methods or the like, and may comprise sub-components which may take the form of instructions or sequences of instructions at any of the levels of abstraction, from the direct machine instructions of a native instruction set to high-level compiled or interpreted language constructs.

Embodiments of the present techniques also provide a non-transitory data carrier carrying code which, when implemented on a processor, causes the processor to carry out any of the methods described herein.

The techniques further provide processor control code to implement the above-described methods, for example on a general purpose computer system or on a digital signal processor (DSP). The techniques also provide a carrier carrying processor control code to, when running, implement any of the above methods, in particular on a non-transitory data carrier. The code may be provided on a carrier such as a disk, a microprocessor, CD- or DVD-ROM, programmed memory such as non-volatile memory (e.g. Flash) or read-only memory (firmware), or on a data carrier such as an optical or electrical signal carrier. Code (and/or data) to implement embodiments of the techniques described herein may comprise source, object or executable code in a conventional programming language (interpreted or compiled) such as Python, C, or assembly code, code for setting up or controlling an ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array), or code for a hardware description language such as Verilog® or VHDL (Very high speed integrated circuit Hardware Description Language). As the skilled person will appreciate, such code and/or data may be distributed between a plurality of coupled components in communication with one another. The techniques may comprise a controller which includes a microprocessor, working memory and program memory coupled to one or more of the components of the system.

It will also be clear to one of skill in the art that all or part of a logical method according to embodiments of the present techniques may suitably be embodied in a logic apparatus comprising logic elements to perform the steps of the above-described methods, and that such logic elements may comprise components such as logic gates in, for example a programmable logic array or application-specific integrated circuit. Such a logic arrangement may further be embodied in enabling elements for temporarily or permanently establishing logic structures in such an array or circuit using, for example, a virtual hardware descriptor language, which may be stored and transmitted using fixed or transmittable carrier media.

In an embodiment, the present techniques may be realised in the form of a data carrier having functional data thereon, said functional data comprising functional computer data structures to, when loaded into a computer system or network and operated upon thereby, enable said computer system to perform all the steps of the above-described method.

The methods described above may be wholly or partly performed on an apparatus, i.e. an electronic device, using a machine learning or artificial intelligence model. The model may be processed by an artificial intelligence-dedicated processor designed in a hardware structure specified for artificial intelligence model processing. The artificial intelligence model may be obtained by training. Here, “obtained by training” means that a predefined operation rule or artificial intelligence model configured to perform a desired feature (or purpose) is obtained by training a basic artificial intelligence model with multiple pieces of training data by a training algorithm. The artificial intelligence model may include a plurality of neural network layers. Each of the plurality of neural network layers includes a plurality of weight values and performs neural network computation by computation between a result of computation by a previous layer and the plurality of weight values.

As mentioned above, the present techniques may be implemented using an AI model. A function associated with AI may be performed through the non-volatile memory, the volatile memory, and the processor. The processor may include one or a plurality of processors. At this time, one or a plurality of processors may be a general purpose processor, such as a central processing unit (CPU), an application processor (AP), or the like, a graphics-only processing unit such as a graphics processing unit (GPU), a visual processing unit (VPU), and/or an AI-dedicated processor such as a neural processing unit (NPU). The one or a plurality of processors control the processing of the input data in accordance with a predefined operating rule or artificial intelligence (AI) model stored in the non-volatile memory and the volatile memory. The predefined operating rule or artificial intelligence model is provided through training or learning. Here, being provided through learning means that, by applying a learning algorithm to a plurality of learning data, a predefined operating rule or AI model of a desired characteristic is made. The learning may be performed in a device itself in which AI according to an embodiment is performed, and/o may be implemented through a separate server/system.

The AI model may consist of a plurality of neural network layers. Each layer has a plurality of weight values, and performs a layer operation through calculation of a previous layer and an operation of a plurality of weights. Examples of neural networks include, but are not limited to, convolutional neural network (CNN), deep neural network (DNN), recurrent neural network (RNN), restricted Boltzmann Machine (RBM), deep belief network (DBN), bidirectional recurrent deep neural network (BRDNN), generative adversarial networks (GAN), transformers and deep Q-networks.

The learning algorithm is a method for training a predetermined target device (for example, a robot) using a plurality of learning data to cause, allow, or control the target device to make a determination or prediction. Examples of learning algorithms include, but are not limited to, supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning.

BRIEF DESCRIPTION OF THE DRAWINGS

The application file contains at least one drawing executed in color. Copies of this patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

Implementations of the present techniques will now be described, by way of example only, with reference to the accompanying drawings, in which:

FIG. 1A is a schematic representation of a machine learning model which may be used for image classification;

FIG. 1B is a schematic representation of class incremental learning (CIL) applied to the model of FIG. 1A;

FIG. 1C is a table showing the definitions of various probability terms used in the methods;

FIG. 2A is a flowchart of the method of the present techniques;

FIG. 2B is an example algorithm for implementing the method of FIG. 2A;

FIG. 2C schematically illustrates modelling semantic drift;

FIG. 2D schematically illustrates modelling feature drift;

FIGS. 3A and 3C schematically illustrate one ML model which may be used to learn and then infer feature drift, respectively;

FIGS. 3B and 3D schematically illustrate an alternative ML model which may be used to learn and then infer feature drift, respectively;

FIG. 3E is a schematic illustration of a Gaussian model which may be used;

FIGS. 4A and 4C schematically illustrate one ML model which may be used to learn and then infer semantic drift, respectively;

FIGS. 4B and 4D schematically illustrate an alternative ML model which may be used to learn and then infer semantic drift, respectively;

FIG. 5A is a flowchart showing more detail of the training of the semantic and feature drift models in the method;

FIG. 5B is a flowchart showing more detail of the training of the classification model in the method;

FIG. 5C shows a block diagram of an apparatus for training the models and implementing the trained model;

FIGS. 6A to 6C plot the accuracy (Top-1%) against number of steps for experimental results using the Tiny-Imagenet benchmark dataset with 5, 10 and 20 phases being used respectively;

FIGS. 6D to 6F plot the accuracy (Top-1%) against number of steps for experimental results using the CIFAR100 benchmark dataset with 5, 10 and 20 phases being used respectively;

FIG. 6G is a table of the results from FIGS. 6A to 6F;

FIGS. 7A to 7C plot the accuracy (Top-1%) against number of steps for experimental results using the CUB 200-2011 benchmark dataset with 5, 10 and 20 phases being used respectively;

FIG. 7D is a table of the results from FIGS. 7A to 7C;

FIGS. 8A to 8C plot the class-average forgetting measure against number of steps for experimental results using the CUB 200-2011 benchmark dataset with 5, 10 and 20 phases being used respectively;

FIGS. 8D to 8F plot a forgetting measure against number of steps for experimental results using the CUB 200-2011 benchmark dataset with 5, 10 and 20 phases being used respectively;

FIGS. 9 and 10 plot the feature representation of samples of the first four learned classes for the PASS method and the proposed method respectively;

FIG. 11 is a feature representation of samples of the first four learned classes for the proposed method applied to the CIFAR 100 benchmark dataset using 20 incremental steps;

FIGS. 12A and 12B are feature representations of samples of the first four learned classes for the proposed method applied to the TinyImageNet and CUB 200-2011 benchmark datasets using 20 incremental steps, respectively;

FIGS. 13A and 13B plot the average Euclidean distance between estimated prototypes of old classes and their true representations and the average cosine similarity between estimated prototypes of old classes and their true representations against incremental step respectively for each of PASS, SDC and the method proposed above applied to the CIFAR 100 benchmark dataset;

FIGS. 13C and 13D plot the average Euclidean distance between estimated prototypes of old classes and their true representations and the average cosine similarity between estimated prototypes of old classes and their true representations against incremental step respectively for each of PASS, SDC and the method proposed above applied to the TinyImageNet benchmark dataset;

FIGS. 13E and 13F plot the average Euclidean distance between estimated prototypes of old classes and their true representations and the average cosine similarity between estimated prototypes of old classes and their true representations against incremental step respectively for each of PASS, SDC and the method proposed above applied to the CUB 200-2011 benchmark dataset;

FIGS. 13G, 13I and 13K plot the average Euclidean distance between estimated prototypes of old classes and the evanescent prototypes of old classes against number of classes for each of PASS [43], SDC [40] and the method proposed above applied to the CIFAR100, TinyImageNet and CUB 200-2011 benchmark datasets respectively using a MLP;

FIGS. 13H, 13J and 13L plot the average Euclidean distance between estimated prototypes of old classes and the evanescent prototypes of old classes against number of classes for each of PASS [43], SDC [40] and the method proposed above applied to the CIFAR100, TinyImageNet and CUB 200-2011 benchmark datasets respectively using a VAE;

FIGS. 14A and 14D plot the pairwise normalised Euclidean distance between prototypical feature representations of classes introduced in the first incremental phase for the PASS method and our method respectively;

FIGS. 14B and 14E plot the pairwise normalised Euclidean distance between prototypical feature representations of classes introduced in the second incremental phase for the PASS method and our method respectively;

FIGS. 14C and 14F plot the absolute values of the difference for the PASS method and our method respectively;

FIGS. 15A to 15C plots the top-1 accuracy (%) in relation to normalised prototype Euclidean distance to reference (ground-truth) representations for each of our method, PASS and SDC methods applied to the CIFAR100, TinyImageNet and CUB 200-2011 benchmark datasets respectively;

FIG. 16A plots the average entropy of the distribution against incremental step for each of PASS, SDC and our method;

FIG. 16B plots the cross-entropy of the distribution against incremental step for each of PASS, SDC and our method;

FIGS. 17A and 17E plot the class-wise entropy of the feature distribution against incremental step for our method and PASS applied to the CIFAR100 dataset, respectively;

FIGS. 17B and 17F plot the class-wise entropy of the class distribution against incremental step for our method and PASS applied to the CIFAR100 dataset, respectively

FIGS. 17C and 17G plot the cross-entropy of the class distribution against incremental step for our method and PASS applied to the CIFAR100 dataset, respectively;

FIGS. 17D and 17H plot the class-wise top-1 accuracy against incremental step for our method and PASS, respectively applied to the CIFAR100 dataset;

FIGS. 18A and 18B plot the semantic drift loss against optimisation stage for the first task and the fifth task when using the MLP adaptation on the CUB 200-2011 dataset;

FIGS. 18C and 18D plot the semantic drift loss against optimisation stage for the first task and the fifth task when using the VAE adaptation on the CUB 200-2011 dataset;

FIGS. 18E and 18F plot the feature drift loss against optimisation stage for the first task and the fifth task when using the MLP adaptation on the CUB 200-2011 dataset; and

FIGS. 18G and 18H plot the feature drift against optimisation stage for the first task and the fifth task when using the VAE adaptation on the CUB 200-2011 dataset.

DETAILED DESCRIPTION

As background, it is useful to understand the terms concept drift and representation drift.

Concept drift refers to the change of the input-output relationship (statistics) undergone by data seen by a ML framework. When the ML model is trained on a dataset, it might suffer from performance degradation if applied to data with distribution evolving over time. Concept drift does not necessarily involve a change in model parameters but focuses on how data distribution has drifted from each training phase of the ML model.

Representation drift indicates how feature representations learned by a ML model are constantly evolving throughout its training procedure. When optimizing a ML model with training samples, model parameters are updated to provide feature representations that better help to solve the target task on the available samples. In a continual learning framework, the constantly changing feature representations used to better model current training data distribution leads to catastrophic forgetting as illustrated in FIG. 1B. This means that representations extracted from past data are becoming staler and progressively less useful to solve the target task. When modeling the drift in feature representations, we can prevent forgetting by employing updated knowledge of no longer accessible feature representations from former training phases.

As explained in more detail below, the proposed new framework may be considered to learn the relationship and drift between old and new classes in order to re-estimate the semantic representations (e.g. prototypes) of old classes using the continually evolving feature representations. We show that our learnable framework modeling representation drift provides models with higher capacity, and leads to overall improved classification performance. For example, experiments on multiple benchmarks show that the proposed approach achieves SotA results for the CIL problem. The proposed new framework uses CIL and more background on the learning representations for CIL is set out below.

Learning Representations in CIL

At each step t∈[T]={0, 1, . . . , T} of CIL, we are given a dataset _t=(_t,_t), t=0, . . . , T, where _t={x_t,j}_j=1^Nt is the set of samples, _t={y_t,j∈_t}_j=1^Nt is the set of their labels. _t is the set of class labels observed at this step, and _t∩_t′=Ø, ∀t≠t′.

As in other popularly employed CIL models [43], the proposed framework is composed of a feature extraction model f_θ∈ and a classification model h_ϕ∈ with parameters θ∈Θ and ϕ∈Φ. An example of such a CIL model is shown in FIG. 1A. At each t^th step, the models are trained by any suitable technique for example using a loss function such as:

$\begin{array}{cc}\underset{{\theta }_t,{\varphi }_t,ϵ}{\mathrm{argmin}}{L}_{\mathrm{cc}}^t+{\sum }_{t^{\prime }=0}^{t-1}{ϵ}_{t\prime },& \left(1\right)\end{array}$

where L_cc^tL_μt(_t; θ_t, ϕ_t) is the expected loss of the model g_th_ϕtºf_θt on the data _t sampled from a distribution μ_t at step t. The slack variable ϵ_t′≥L_cp,t′−L_pp,t′ measures the difference between the loss of the current model on the previous datasets {_t′}_t′=0^t−1 denoted by L_cp,t′l(g_t(_t′))), and the loss of the previous models {g_t′}_t′=0^t−1 on the previous datasets {_t′}_t′=0^t−1 denoted by L_pp,t′l(g_t′(_t′)). The variable ϵ_t, controls forgetting of representations of old classes.

Generative classifiers, such as Bayesian networks or supervised variational autoencoders (VAEs), implementing h_ϕ∈ optimize equation (1) to model joint distribution of classes and features p(C, F; ) where =Θ∪Φ is the set of parameters of the model. Discriminative classifiers, such as a softmax classifier, optimize equation (1) to model p(C|F; ).

In CIL, we do not have access to samples {_t′}_t′=0^t−1 at time t. CIL methods aim to model p(C, F|_t) without using {_t′}_t′=0^t−1, where

• • _t=Θ_t∪Φ_t, Θ_t={θ_t′}_t′=0^t and Φ_t={ϕ_t′}_t′=0^t−1,
  • F∈=_t∪_old is the random variable taking values from the set of feature representations _t learned at step t on _t and from _old={_t′}_t′=0^t−1, and
  • C∈=_t∪_old is a random variable of semantic (class) representations, where _old={_t′}_t′=0^t−1.It is noted that in some CIL models, models trained at earlier steps t′<t are frozen and reused at consecutive steps, while the other methods incrementally update the models at each step. In the latter case, _t={θ_t}∪{ϕ_t}

In order to elucidate the dynamics of models used for CIL in this setting, we factorize the class posterior probability p(C∈|F∈) by

$\begin{array}{cc}p\left(C\in 𝒞❘F\in ℱ\right)\propto \frac{P_A-P_B}{P_C},& \left(2\right)\end{array}$

where P_C, P_A: =P₁+P₂+P₃+P₄, and P_B: =P₁₂+P₁₃+P₁₄+P₂₄+P₃₄ are expressed in Table 1 shown in FIG. 1C. It is noted that, for the sake of simplicity, _t is dropped from the statements when dependency on _t is trivial.

Deep learning models have been employed to estimate and model these probabilities implicitly at different parts of the models by learning feature representations of old and new classes, and make predictions for the new classes as follows:

• • Class posterior probabilities for old (p(C∈_old|•)) and new classes (p(C∈_t|•)) are computed using classifiers h_Φt′∀t′∈[t].
  • Feature representations _t of new classes _t are learned by updating f_Θt′, ∀t′∈[t].Evanescent Representations: At the t^th step, we do not have access to old samples _old={_t′}_t′=0^t−1. Thus, features _old cannot be extracted from _old. Therefore, state-of-the-art methods aiming to learn representations considering equation (2) treat the term P_C, which denotes distribution of features learned using old classes, as a normalizing partition function, and ignore it to compute

p(C∈|F∈)∝P_A−P_B.  (3)

As explained in more detail below, to address this problem and bring the evanescent representations to life, we employ class prototypes (templates or prototypical representations, the terms can be used interchangeably) π∈Π_old. These class prototypes can be obtained using any standard technique, for example as taught in [43] and may be regarded as semantic representations of C_old. We leverage prototypes at the beginning of an incremental step to model their distribution p(F∈_old|π_c∈Π_old) where π_c which is the class prototype for the class c. We estimate the distribution as described in more detail below, then we update p(F∈_old) throughout the incremental step by modeling the representation drift to revive evanescent representations.

FIG. 2A is a schematic flowchart of the method which is used to overcome the problems outlined above and which is described in more detail below. In general terms this new framework models the drift undergone by data representations by (i) learning the semantic drift between past and novel classes and (ii) estimating the evolution of features extracted by the classification model. Semantic and feature drifts may be jointly exploited to access up-to-date representations of data of past classes. Merely as an example, in addition to the improvements noted over [43] in the background section, [43] also does not directly capture the relationship between semantic representations of old (Π_old) and new classes (Π_t′), rather focusing on estimating the connection between across-task drifts of the representations, and neglects what we call feature representation drift, that is the joint evolution of model features during single task training (i.e., they treat feature representation drift independently for each feature channel).

FIG. 2B shows an example of an algorithm which can be used to implement the training of the classifier as shown in the flowchart of FIG. 2A. The first step of the algorithm is to input a dataset {_t=(_t,_t)}_t=0^T where _t={x_t,j}_j=1^Nt is the set of samples, _t={y_t,j∈_t}_j=1^Nt is the set of their labels. _t is the set of class labels observed at this step, N_t is the number of samples per step and _t∩_t′=Ø, ∀t≠t′. This first step is shown as S200 in FIG. 2A and at this first step, in addition to the dataset, the classifier which is to be trained is also obtained. The proposed framework is composed of a feature extraction model f_θ∈ and a classification model h_ϕ∈.

As shown in the algorithm and also in step S202, the next step is to train both the feature extraction model f_θo and the classification model h_ϕ0 for the incremental time step 0 using a loss function L_cc⁰ on the initial dataset ₀. Once the model has been trained for time step 0, the class prototype(s) Π_new⁰ can be computed and then we initialise Π_old^t=1,0 as Π_new⁰. The loss function and the computation of the class prototypes are described in more detail below. Steps S202 and S204 may thus be considered an initialisation phase.

The following processes are then repeated for each step t>0. As shown at step S206, the set of features at the beginning of the step _t⁰ and at the nth optimisation stage of the step _tⁿ are extracted or computed using the trained feature extractor model and the new class data _t for that time step. As shown in steps S208 and S210, the relationship between old and new semantic and feature representations are modeled using two models. A feature drift model Γ_γ is parameterized by γ, and is used to represent feature drift. A semantic drift model Ψ_ψ is parameterized by ψ and is used to represent semantic drift. The parameters γ and ψ may be optimized in n epochs (or stages) using an optimization algorithm, while training models at each incremental step to estimate the relationship among representations. This first phase may be termed a learning phase, and the models Γ_γtn and Ψ_ψtn at each stage n and each step t are trained by any suitable method either jointly or separately as described in more detail below.

Once the models Γ_γ and Ψ_ψ have optimized parameters, they are used to estimate the revived evanescent representations (RERs) as shown at step S212. The method of estimating the revived evanescent representations depends on the architecture as explained in more detail below. For example, the revived evanescent representations may be estimated by first estimating _oldⁿ which are the set of features for the revived evanescent representations at the nth optimisation stage. As explained in more detail below, _oldⁿ may be estimated by using the semantic drift model and feature drift model individually or fused. Then, the old class prototypes Π_old^t,n may be computed by class-wise averaging features sampled from p(F∈_oldⁿ). This phase may be termed the inference phase.

The next step S214 is then to train the feature extraction model f_θt and the classification model h_ϕt on the updated dataset which is the original dataset for the time step together with the revived evanescent representation, i.e. _t=_t∪_oldⁿ. As for the semantic and feature drift models, the parameters of the feature extraction and classification models may be optimized in n epochs (or stages) using an optimization algorithm. The training may be done using any suitable technique as described in more detail below.

As shown in steps S208, S210 and S214, there is alternating training of the semantic drift and feature drift models, and the feature extraction and classification models for each optimisation stage. As shown in step S216, there is an inner loop within the algorithm to include an optional optimisation at each incremental time step t. The optimisation may be over N stages (also termed epochs). If n<N, there is an iteration through steps S206 to S214 to train the semantic drift and feature drift models Γ_γtn and Ψ_ψtn in the first phase (learning phase) and then in the second phase (inference phase) to infer the revived evanescent representations which are used to train the feature extraction model f_θt and the classification model h_ϕt.

In other words, at the end of each optimization stage n, the feature extraction model f_θt and the classification model h_ϕt are updated (step 7 of Algorithm 1). It is noted that at n=0, the feature extraction model f_θt and the classification model h_ϕt are not updated either in the learning phase or the inference phase because they are already known either from the previous time incremental step or from the initialisation. After the inference phase for n=0, we update the feature extraction model f_θt and the classification model h_ϕt. These updated models are used in the learning and inference phases for n=1. This procedure is then implemented at each stage.

Once there are no more stages, i.e. for n=N, as shown at step S216, there is a final learning phase for the incremental step in which the semantic drift and feature drift models Γ_γtN and Ψ_ψtN are trained at steps S218 and S220. Once all the optimisation stages have been completed, at step S222, we compute the new prototypes Π_new^t={π_c, c∈_t} by class-wise average of feature representations f_θt(_t) of input samples and re-initialise the complete set of class prototypes for the next time step as Π_old^t+1,0=Π_old^t,n∪Π_new^t, where Π_old^t,n=Ø for t=0. In other words, there is another inference phase to obtain the set of class prototypes which contains the newly obtained class prototypes and the previously obtained set of old class prototypes. As shown at step S224, if there are more time steps there is an iteration through steps S206 to S220 to train the semantic drift and feature drift models Ψ_ψtN and Γ_γtN in the first phase (learning phase) and then in the second phase to infer the revived evanescent representations which are used to train the feature extraction model f_θt and the classification model h_ϕt. Optimisation stages can also be employed at the time step as described above.

FIG. 2C schematically illustrates the semantic drift which is modelled using the semantic drift model. Merely as an example, three classes are shown in FIG. 2C and these may be same as the three classes described in relation to FIG. 1B. At optimisation stage 0, several items or images which are schematically indicated by triangles have the same classification, e.g. an image of a dog. The images form a cluster 112 which is centred on a prototype image 114 and there is a decision boundary 110 of the feature space within which all images would be classified as a dog. There are also several images which are schematically indicated by the squares and which may be images of zebras and several images which are schematically indicated by the circles and which may be images of horses The decision boundaries 120, 130 and prototypes 124, 134 for these classes are also illustrated.

As shown in FIG. 2C after n optimisation stages, the decision boundaries 120′, 130′ and prototypes 124′, 134′ for these second and third classes (e.g. zebras and the horses) have been optimised. Thus, they are in a different location within the feature space. FIG. 2C also shows the revived evanescent representations of the images of the dogs which are not included in the dataset which was used in this time step. The revived evanescent representations 118 are indicated by partially filled triangles. A revived class prototype 114′ is also determined for the revived evanescent representations 118. The semantic drift models the relationship between the prototypes of each class and in particular how this changes over the optimisation stages. This is indicated by the arrows connecting the three different prototypes.

FIG. 2D schematically illustrates the feature drift which is modelled using the feature drift model. As with FIG. 2C, three classes are shown and the same numbering is used. The new classes are shown on the left part of FIG. 2D and here the evolution of the features for each class is modelled separately. Thus, the arrows in this instance connect the original prototype for each class with the new prototype at the end of the optimisation stages. Similarly, the old class is shown on the right of the image and the arrows connect the original prototype for this class with the estimated prototype generated from the revived evanescent representations.

In other words, the feature learning model may be represented by f_new,tⁱ which is an evolving feature representation on a sample X_newⁱ from a new class. Although the example given above is in the field of image classification, the sample may be an image, text data or audio data. As shown in FIGS. 2C and 2D, there is modelling of both feature drift and semantic drift for semantic representations (e.g. prototypes) of old classes. The feature drift may be learnt from the available data as denoted by:

{F_new,0ⁱ}_i↔{F_new,tⁱ}

An updated prototype (also termed semantic representation) for the revived evanescent representations which is learnt from this modelling may be expressed as:

{π_old,t^j}_j^{tr j}

The semantic drift is learnt from the available prototypes as denoted by:

{F_new,0ⁱ}_i↔{π_old,0^j}_j

The updated prototype for the revived evanescent representations which is learnt from this modelling may be expressed as:

{π_old,t^j}_j^cls

The updated prototype for the revived evanescent representations which is learnt from both types of modelling may converge to a unique representation which is expressed as:

{π_old,0^j}_j

The following subsections present the method proposed for modelling drifts. To identify and train the feature drift and semantic drift models Γ_γ and Ψ_ψ, two different core architectures are proposed. A first core architecture which may be used is a Gaussian model (GM) in which the class conditional feature distribution is modelled as a parametric Gaussian curve. The second core architecture which may be used is a variational model (VM) which comprises a conditional variational encoder (VAE) having an encoder and a decoder separated by a latent space. Such GM and VM models are parametrized by deep neural networks (DNNs). Although Gaussian processes can be used for VMs [14], we consider GMs and VMs individually to explicate the variational structure of VMs.

Modelling Feature Drift

FIGS. 3A to 3D show two types of architecture which may be used to model feature drift. In a first variation shown in FIGS. 3A and 3C, we use a deep neural network (DNN) in the form of a simple, linear and lightweight multilayer perceptron (MLP) to implement the model which maps the set of features extracted at the beginning of the incremental step t to the set of features at the nth optimisation stage of step t, i.e. Γ_γtn: _t⁰→_tⁿ. As shown in FIGS. 3A and 3C, the MLP is composed of two fully-connected (FC) layers and a ReLU activation between them. The input and output variables of Γ_γtn correspond to sets of B feature vectors of dimension D (number of feature channels at the output of the feature extractor, which is set to 512 in all experiments), both arranged in a B×D matrix. In addition, the number of output channels of the first FC layer is set to 2*D and B is set equal to the cardinality of _t (i.e., the available training set at step t). The architecture may be summarised in the table below.

Feature drift
	Input	Operator	Output
	B × D	FC layer	B × 2D
	B × 2D	FC layer	B × D

In a second variation shown in FIGS. 3B and 3D, we use a nonlinear lightweight conditional variational auto-encoder [for example as described in 11] to implement Γ_γtn. The task of the cVAE is to learn a generative function of feature representations of training samples at stage n (_tⁿ), conditioned on the representations of the same samples at the beginning of the current incremental step (_t⁰). The encoder and decoders are composed of two FC layers each. We perform conditioning in input and latent spaces by concatenation along the channel dimension. The input, output and conditioning variables of Γ_γtn correspond to sets of B feature vectors of dimension D, all arranged in B×D matrices. We set B equal to the cardinality of _t and C to the cardinality of C_old^t. The architecture may be summarised in the table below.

Feature drift
Encoder	Decoder
Input	Operator	Output	Input	Operator	Output
B × 2D	FC layer	B × 4D	B × 2D	FC layer	B × 4D
B × 4D	FC layer	B × 2D	B × 4D	FC layer	B × 2D

FIGS. 3A and 3B are used in the learning phase which is described above in the flowchart. In this learning phase, we aim at modelling the feature drift (FD) on the dataset _t as representations revive and evolve throughout the step t>0. At stage n>0, we extract the set of features at the beginning of the time step _t⁰ and the set of features extracted after n>0 optimisation stages to train the model Γ_γtn to learn the relationship between _t⁰ and _tⁿ. Both _t⁰ and _tⁿ are thus part of the available information on the new class features. The model weights (also known as parameters γ) for the first model Γ_γ are updated in this first phase.

When using a GM as shown in FIG. 3A, we first identify Γ_γtn: _t⁰→_tⁿ by the MLP. Then, Γ_γtn is trained to track and model the evolution of revived evanescent representations (RERs) from stage 0 to n. The model is learned using any suitable technique as described in more detail below.

When using a VM as shown in FIG. 3B, we consider Γ_γtn as a stochastic map and identify it by a variational model such as a variational auto-encoder (VAE). The VM is trained by maximising the likelihood p(F∈_tⁿ|F∈_t⁰; γ_tⁿ). Thereby, we can statistically model the FD across different stages [n] at a given step t.

FIGS. 3C and 3D are used in the interference phase (or second phase) which is described above in the flowchart. In the first phase, the model Γ_γ is trained to learn the drift experienced by the classes available at the current step _t. In this second phase, we exploit the trained model to infer the feature drift undergone by features of _old and distribution p(F∈_oldⁿ) of revived evanescent representations.

As shown in FIG. 3C, the DNN is fixed with the model Γ_γtn. The model is then used to infer representations of old class prototypes Π_old^t,n at stage n>0 using representations of old class prototypes at the beginning of the time step Π_old^t,0 as the input. In other words, the model enables to directly track the trajectory of the representations of old class prototypes from stage 0 to n and thus map Π_old^t,0 to Π_old,f^t,n. If n=0, in other words, if no optimisation is used, then Γ_γto is an identity map. Finally, we approximate the distribution p(F∈_oldⁿ) of revived evanescent representations at stage n by a Gaussian distribution which is schematically illustrated in FIG. 3E. The Gaussian distribution may be expressed as p(F∈_oldⁿ: π_c∈Π_old,f^t,n)˜N(π_c,σ_c), where N is the normal distribution and σ_c is the standard deviation estimated at step t′ when c∈C_t′. As illustrated in FIG. 3E, the standard deviation is kept fixed at every step t″>t′.

As shown in FIG. 3D, the encoder is not used and the decoder is kept fixed in the inference phase. The trained model Γ_γtn is used to provide an approximation of the probability that the set of features is a member of the set of features for the revived evanescent representations at the nth optimisation stage given the set of features is a member of the set of features for the revived evanescent representations at the start of the time step. p(F∈_oldⁿ|F∈_old⁰). At stage n=0, we resort to p(F∈_old⁰) because no feature drift has to be estimated, and we model the distribution of revived evanescent representations by p(F∈_old⁰))∝p(F∈_oldⁿ: π_c├Π_old^t,0)˜N(π_c,σ_c), that is, the distribution p(F∈_old⁰)) is approximated by the empirical model p(F∈_oldⁿ: π_c∈Π_old^t,0) identified by prototypes π_c∈Π_old^t,0 which is estimated by a Gaussian model N(π_c,σ_c) with mean set to the prototype π_c and with the standard deviation σ_c. At n>0, the training features are sampled from the posterior probability p(F∈_oldⁿ|F∈_old⁰; γ_tⁿ)·p(F∈_old⁰). The standard deviation is kept fixed at every step t″>t′. The input of the model is the prototype set Π_old^t,0 and noise z˜N(0, I) is drawn from the Normal distribution with zero mean and unit variance. The output of the model is the set of features for the revived evanescent representations, i.e. _oldⁿ.

In other words, in the second phase the prototypes of the old classes are modelled using the model learnt in the first phase. The model Γ_γ is thus implemented in the second phrase which uses a probabilistic framework to estimate feature drift on the unavailable past data. The model weights for the first model Γ_γ are kept fixed in phase 2. In this example, in the first phase, the inputs of the model are the feature set _t⁰ and the feature set _tⁿ and both of these inputs are available. The output of the model is the set of reconstructed features for the nth stage _tⁿ. In the second phase, the input of the model is the prototype set Π_old^t,n=0 which is available data and the output of the model is the feature set for the revived evanescent representations and/or the old prototypes. Both of these outputs are inferred by the model.

Modelling Semantic Drift

FIGS. 4A to 4D show two types of architectures which may be used to model semantic drift. In a first variation shown in FIGS. 4A and 4C, we use a DNN in the form of a simple and lightweight multilayer perceptron (MLP) to implement the model which maps the set of features extracted at the beginning of the incremental step t to the set of features at the nth optimisation stage of step t, i.e. Ψ_ψtn: _n⁰→Π_old^t,0. As shown in FIGS. 4A and 4C, the MLP is composed of two fully-connected (FC) layers and a ReLU activation between them. The input and output variables of Ψ_ψtn correspond to sets of B and _old^t feature vectors of dimension D both arranged in a D×B and a D×_old^t matrix, respectively. _old^t denotes the number of past classes present at the current incremental step t>0 and C is set equal to the cardinality of _old^t. D is the number of feature channels at the output of the feature extractor (which is set to 512). In addition, the number of output channels of the first FC layer is set to 2*B and B is set equal to the cardinality of _t (i.e., the available training set at step t). The architecture may be summarised in the table below.

Semantic drift
	Input	Operator	Output
	D × B	FC layer	D × 2B
	D × 2B	FC layer	D × C

In a second variation shown in FIGS. 4B and 4D, we use a lightweight conditional variational auto-encoder (cVAE) to implement Ψ_ψtn. The task of the cVAE is to learn a generative function of feature representations of old classes (_old⁰) whose distribution is approximated as p(_old⁰; Π_old^t,0)˜N(π_c,σ_c). The function is conditioned on the new classes _t⁰ which can be extracted from the dataset. The encoder and decoder are composed of two FC layers each. We perform conditioning in input and latent spaces by concatenation along the channel dimension. The input, output and conditioning variables of Ψ_ψtn correspond to sets of C_old^t feature vectors of dimension D, all arranged in D×C_old^t matrices. The conditioning variable is a set of B feature vectors of dimension arranged in a D×B matrix. We set B equal to the cardinality of _t and C to the cardinality of C_old^t. The architecture may be summarised in the table below.

Semantic drift
Encoder	Decoder
Input	Operator	Output	Input	Operator	Output
D × (C + B)	FC layer	D × 2(C + B)	D × (1 + B)	FC layer	D × 2(1 + B)
D × (2C + B)	FC layer	D × 2	D × 2(1 + B)	FC layer	D × C

FIGS. 4A to 4D shows more detail of the two phases which are used when modelling semantic drift using the second model Ψ_ψ (also termed a semantic drift model). FIGS. 4A and 4B are used in the learning phase which is also known as the first phase. In this phase, we aim at capturing the semantic drift (SD) experienced by representations at each incremental step. To this end, at the beginning of each step, we extract the set of features at the beginning of the incremental step _t⁰ and train a network or semantic drift model Ψ_ψot with parameters ψ₀^t to model the relationship between the set of features at the beginning of the incremental step _t⁰ and the set of features of the evanescent representations at the beginning of the time step _old⁰. We employ the prototypes π∈Π_old^t,0 to model p(F∈_old⁰))∝p(F∈_oldⁿ: π_c∈Π_old^t,0)˜N(π_c,σ_c). The old class prototypes Π_old^t,0 at stage 0 are known. As opposed to the feature drift model, the semantic drift model captures the semantic drift observed at each new step and thus an individual model Ψ_ψ is optimised at the start, and fixed for the rest of the step (i.e. ψ₀ⁿ=ψ₀^t). As explained in more detail below, when model fusion is adopted, instead the semantic drift model Ψ_ψ is re-trained once per stage, to account for the feature drift estimated by the feature drift model.

When using a GM as shown in FIG. 4A, we first identify the semantic drift model using a DNN which maps the set of features at the beginning of the incremental step _t⁰ to the old prototypes at the beginning of the incremental step. In other words, we identify Ψ_ψtn: _t⁰→Π_old^t,0. Then the semantic drift model is trained to model the semantic drift between representations for classes available at the current step _t and those experienced in the past _old. The model is learned using any suitable technique as explained in more detail below.

In the non-linear implementation shown in FIG. 4B, we first approximate the distribution p(F∈_oldⁿ) of evanescent representations revived at stage n=0 by p(F∈_oldⁿ: π_c∈Π_old^t,0)˜N(π_c,σ_c). Then a conditional VM (e.g. a VAE) is trained by adjusting the weights to maximise the probability that a set of features are within the features of evanescent representations revived at stage n>0 given that the set of features are within the features of the new class features at stage 0, i.e. maximising the likelihood p(F∈_oldⁿ|F∈_t⁰; γ). The training can be done using any suitable technique.

FIGS. 4C and 4D are used in the interference phase (or second phase). In the learning phase, the semantic drift model Ψ_ψtn is trained to learn the SD experienced while moving from the representations for old classes _old to the representations for the current classes _t. The drift is captured at the beginning of the current step t, when up-to-date representations of both sets are available. We now exploit the trained semantic drift model Ψ_ψtn to infer the semantic drift undergone at every stage n>0, and estimate the distribution of the probability that the set of features for each revived evanescent representation is the old class features, i.e. p(F∈_oldⁿ).

FIG. 4C shows the linear implementation and shows that the trained model Ψ_ψtn is used to infer the representations of old class prototypes Π_old^t,n∀n≥0 under semantic drift. That is, the model Ψ_ψtn: _tⁿ→Π_old^t,n is trained to estimate the relationship between feature and prototypical representations at stage n. In a similar manner to that described for FIG. 3C, we then approximate the distribution p(F∈_oldⁿ) of revived evanescent representations at stage n by a Gaussian distribution p(F∈_oldⁿ: π_c∈Π_old,s^t,0)˜N(π_c,σ_c) where N is the normal distribution and σ_c is the standard deviation estimated at step t′ when c∈C_t′.

As shown in FIG. 4D, the encoder is not used and the decoder is kept fixed in the inference phase. The trained model Ψ_ψtn is used to provide an approximation of the probability that the set of features is a member of the set of features for the revived evanescent representations at the nth optimisation stage given the set of features is a member of the set of features for the revived evanescent representations at the nth optimisation stage. p(F∈_oldⁿ|F∈_tⁿ)∀n≥0. To generate training feature samples, we perform inference using p(F∈_oldⁿ|F∈_old⁰; ψ_tⁿ)·p(F∈_tⁿ), where the feature set _tⁿ is provided by the feature extraction model f_θt trained for n optimisation stages and applied on samples from the dataset _t.

Thus, as shown in FIGS. 4A and 4C, a linear model such as a perceptron may be used in both phases. In this example, in the first phase, the input of the second model Ψ_ψ⋅ is the feature set _t⁰ and the output of the model is the prototype set Π_old^t,n=0. In the first phase, the model weights for the second model Ψ_ψ⋅ are updated using this input and output. In other words, the semantic drift is learnt. In the second phase, the model Ψ_ψ is implemented. The input of the model is the feature set _tⁿ and the output of the model is the prototype set Π_old^t,n>0 which is inferred using the model. The model weights for the first model Γ_γ are fixed during the second phase.

Thus, as shown in FIGS. 4B and 4D, a nonlinear model such as a variational auto-encoder (VAE) which includes an encoder and a decoder may be used in both phases. In this example, in the first phase, the inputs of the model are the feature set _tⁿ⁼⁰ and the prototype set Π_old^t,n=0. The output of the model is the set of reconstructed prototypes Π_old^t,n=0. In the first phase, the model weights for the second model Ψ_ψ are updated using these inputs and output. In other words, the semantic drift is learnt. In the second phase, the second model Ψ_ψ⋅ is implemented. The input of the model is the feature set _tⁿ and noise z˜N(0, I) is drawn from the Normal distribution with zero mean and unit variance. The output of the model is the old class features _oldⁿ which are inferred by the second model. The model weights for the second model Ψ_ψ are fixed during the second phase. For this semantic branch we use a conditional variational auto-encoder but it is noted that in FIGS. 3B and 3D, we may employ the infoVAE architecture, since we notice improvements in model stability.

Training Details

FIG. 5A is a flowchart detailing the steps in each incremental step t of the learning phase in which the representation drift models are trained using the frozen classification models from the previous time step t. As explained previously, we do not have access to old samples, and so the old features _old⁰ are not known but the old class prototypes at stage 0 are known and denoted by Π_old^t,0. FIG. 5A shows the method includes, at step S500, obtaining the old class prototypes Π_old^t,0 and using them to approximate _old⁰ at beginning of time step t. These are used to train the semantic drift model as described below.

A new class dataset _t is obtained or input at step S502. In step S504, the trained feature extraction model f_θt-1 which was updated at the previous incremental step t is used to extract the feature representations at the beginning of the time step, i.e. _t⁰. Similarly, in step S506, the feature representations at the nth optimisation stage of the time step, i.e. _tⁿ, are extracted using the feature extraction model f_θt for that nth optimisation stage. Both the feature extraction models which are used to extract the feature representations are frozen.

As shown in step S508, the feature drift model Γ_γtn is learned using any suitable technique by using the extracted feature representations. For example, for the GM shown in FIG. 3A, the training may be done by minimising an error, e.g. the mean squared error between the actual set of features at optimisation stage n>0 and the set of features at the same optimisation stage as predicted by the feature drift model given the set of features at the beginning of the time step, i.e. the difference between Γ_γtn(F_t^o) and _tⁿ. The loss function L_f^t for feature drift may be expressed as:

L _f ^t(F_t^o,F_tⁿ;γ_tⁿ)=L_mse,f^t∥Γ_γtn(F_t^o)−F_tⁿ∥₂²

Similarly, the training of the VM in FIG. 3B can be done using any suitable technique, e.g. the objective proposed in [14]. The learning objective may thus be composed of a reconstruction constraint and a regularisation term measuring the KL divergence between the posterior distribution modelled by the encoder and the standard normal prior, plus an additional term to maximise the mutual information between input and latent variables. The loss function L_f^t for feature drift may be expressed as:

L _f ^t(F_t^o,F_tⁿ;γ_tⁿ)=βL_rec,f^t+(1−α)L_kl,f^t+(α−λ_info−1)L_info,f^t

where L_rec,f^t is the reconstruction loss, L_kl,f^t is the KL divergence loss and L_info,f^t is the loss of the InfoVAE, α, β and λ_info are constants, with β set to le1 in all experiments, λ_info set to le1 and α set to −le1 for the CIFAR100 dataset, λ_info set to le2 and α set to −le2 for the test set TinyImageNet and CUB200 datasets.

As shown in step S510, the semantic drift model Ψ_ψtn is learned using any suitable technique by using the extracted feature representations at the beginning of the incremental step and the old class prototypes. For example, for the GM as shown in FIG. 4A, the semantic drift model may be trained by minimising an error, e.g. the mean squared error between the actual set of prototypes at optimisation stage n>0 and the set of prototypes at the same optimisation stage as predicted by the model given the set of features at the beginning of the time step, i.e. the difference between Ψ_ψtn(F_t^o) and Π_old^t,o. The loss function L_s^t for semantic drift may be expressed as:

L _s ^t(F_t^o,Π_old^t,o;ψ_tⁿ)=L_mse,s^t=∥Ψ_ψtn(F_t^o)−Π_old^t,o∥₂²

In the non-linear implementation shown in FIG. 4B, the training may be done by optimizing the variational model (conditional VAE) by maximizing the Evidence Lower Bound (ELBO) as proposed in [5]. The learning objective may thus be composed of a reconstruction constraint and a regularisation term measuring the KL divergence between the posterior distribution modelled by the encoder and the standard normal prior. The loss function L_s^t for semantic drift may be expressed as:

L _s(F_t^o,Π_old^t,o;ψ_tⁿ)==L_rec,s+λ_kldL_kld,s

where L_rec,s^t is the reconstruction loss, L_kl,s^t is the KL divergence loss and λ_kld,s is a constant set to 1 in all experiments, i.e. for the CIFAR100, TinyImageNet and CUB200 datasets.

At step S512, there is an optional step of fusing the feature and semantic drift models. The fusion may be done by joint training or as explained in relation to the inference step by fusing the output of separately trained models. In the example of FIG. 5A, the fusion is done during training. For example, the loss functions L_f^tL_f(F_t⁰, F_tⁿ; γ_tⁿ) and L_s^tL_s(F_t⁰, Π_old^t,0; ψ_tⁿ) may be used to denote the objectives used to individually train feature and semantic drift models Γ_γtn and Ψ_ψtn, respectively. As described above, the exact form of the aforementioned objectives depends on the employed network architecture identifying Γ_γtn and Ψ_ψtn.

The outputs of Γ_γtn and Ψ_ψtn may be fused by jointly training them to estimate p(F∈F_oldⁿ) using GM and VM. For this purpose, we optimise model parameters by minimising a measure of discrepancy between the estimated distributions p(F∈F_oldⁿ; γ_tⁿ) and p(F∈F_oldⁿ; ψ_tⁿ) employing a suitable training objective also termed a fusion loss may be expressed as

L _fus ^t=∥Π_old,s^t,n−Π_old,f^t,n∥₂²+λ_corr∥ρ(Π_old,s^t,n)−ρ(Π_old,f^t,n)∥₂²⁾

where the subscript s and f denotes the updated prototypes of old classes estimated by the semantic and feature drift models, respectively, ∥·∥₂² is the squared ₂ norm, λ_corr>0 is the regularisation parameter and ρ(Π) is the normalised correlation matrix of Π [3]. Finally, the renovated distributions p(F∈F_oldⁿ; γ_tⁿ) and p(F∈F_oldⁿ; ψ_tⁿ) are linearly combined with equal weights to obtain p(F∈F_old^nd).

Then, the overall objective used to learn representation drift may be is defined by:

L _drift ^t =L _s ^t +L _f ^t+λ_fusL_fus^t

where λ_fus>0 is the loss balancing parameter. The values of λ_fus and λ_corr may be experimentally finetuned for each drift model configuration, dataset and incremental set-up as explained in more detail below. In particular, we perform gridsearch such that λ_fus,λ_corr∈{le2, le1, le0, le-1, le-2, le-3, le-4, le-5} and select the best value combination. We note that L_drift^t measures the loss of current models on inferred representations of old classes. Thereby, we aim at reducing forgetting (ϵ_t), i.e. the discrepancy between RERs as estimated by drift models and their evanescent (unavailable) counterparts by training models optimising L_drift^t.

In all the aforementioned setups, we may employ the Adam optimiser, for example as described in [4] with fixed learning rate η and train until convergence by performing early stopping, that is the model is trained until the loss function does not change for a predefined constant number of steps τ=25. We may also experimentally finetune the value of the learning rate η∈{le-3, le-4, le-5} for each drift model configuration, dataset and incremental setup. Finally, we may also apply weight normalisation, for example as described in [10], to Ψ_ψtn for both GM and VM implementations. We may also apply spectral normalisation, for example as described in [8], to weights of Γ_γtn for both GM and VM implementations. The normalisation may result in an improvement in robustness of training convergence.

FIG. 5B is a flowchart detailing the steps in training the classification models using the results inferred from the trained representation drift models. As explained previously, we do not have access to old samples, and so the old features _old⁰ are not known but the old class prototypes at stage 0 are known and denoted by Π_old^t,0. Like FIG. 5A, FIG. 5B shows the method includes, at step S600, obtaining the old class prototypes Π_old^t,0 and using them to approximate _old⁰ at beginning of time step t. These are used to infer the new features using the feature drift model as described below.

The new class dataset _t is obtained or input at step S602. In step S604, the trained feature extraction model f_θt-1 which was updated at the previous incremental step t is frozen and is used to extract the feature representations at the beginning of the time step, i.e. _t⁰. In step S606, the feature extraction model f_θt for the incremental step t is trained so that the new features can be obtained, i.e. _tⁿ. We can enhance the training objective by a distillation loss L_fkd^tL_fkd(_t), for example as described in [8], to reduce the entity of representation drift across incremental tasks. The distillation loss L_fkd^t is defined by the ₂ distance between representations extracted from _t using the current feature extraction model f_θt and the feature extraction model from the previous step f_θt-1, the latter inherited from the previous step and kept fixed. Thereby, the distillation loss L_fkd^t approximates the difference between the overall classification objective L_cc^t and the loss L_pc^t of the previous model f_θt-1 on _t. The distillation loss may be expressed as:

$L_{\mathrm{fkd}}\left(D_t\right)=\sum _{x\in D_t}❘❘f_{{\theta }_t}\left(x\right)-f_{{\theta }_{t-1}}\left(x\right){}_2^2$

L_pc^t is not explicitly mentioned in the equation above but provides information regarding shareability of representations among consecutive steps t−1 and t. Therefore, models optimising L_fkd^t can make use of the feature shareability for learning drifts.

The new class features _tⁿ can be used in step S610 to infer the distribution of revived evanescent representations p(F∈_oldⁿ) using the trained model Ψ_ψtn. At step S608, the fixed feature drift model Γ_γtn is used to infer representations of old class prototypes Π_old^t,n at stage n>0 using representations of old class prototypes at the beginning of the time step Π_old^t,0 as the input. At step S612, the outputs of the two drift models may be optionally fused, e.g. by a linear combination with equal weights. The revived evanescent representations may then be generated at step S614.

At step S616, the classification model h_ϕt is then trained using any suitable technique. For example, a cross entropy loss L_ce^tL_ce(_t) may be used. As an example, the cross entropy loss may be a standard function such as that defined in the book “Deep Learning” by Goodfellow et al. published by MIT. When t>0, to mitigate forgetting of previous tasks, we generate features of the representations of the old classes _old, by modeling the drift and estimating the distribution p(F∈_old). The representation drift loss function L_rd^t may be then calculated using

$L_{\mathrm{rd}}^t\stackrel{\Delta }{=}{L}_{\mathrm{rd}}\left({ℱ}_{\mathrm{old}}^{n,t}\right)=\sum _{F\in {ℱ}_{\mathrm{old}}^{n,t}}{y}_F\log h_{\varphi t}\left(F\right)$

where y_F is the one-hot label vector of the class c_F∈_old and _old^n,t is the set of revived evanescent representations sampled from the estimated distribution using the updated prototypes of _old. In other words, F∈F_old^n,t is estimated by modelling representation drift as described above. The loss L_rd^t approximates L_μ(_t; θ_t, ϕ_t)≤ϵ_t′ of g_t on the previous datasets {_t′}_t′=0^t−1 using their inferred representations.

In the flowchart, the feature extraction model and the classification model are shown as being trained at separate steps but it will be appreciated that the models may be trained together. Then, the overall classification objective L_cc^t computed at each step t is:

L _cc ^t(D_t,F_oldⁿ;θ_t,ϕ_t)=L_ce^t+λ_rdL_rd^t+λ_fkdL_fkd^t,

where L_rd^t and L_fkd^t are used only for t>0 with loss balancing parameters λ_rd>0 and λ_fkd>0. An alternative expression is:

$\underset{{\theta }_t,{\varphi }_t}{\mathrm{argmin}}{L}_{\mathrm{cc}}\left(D_t,F_{\mathrm{old}}^n;{\theta }_t,{\varphi }_t\right)$

where θ_t,ϕ_t are the parameters of the current feature extraction model and classification model, respectively.

In summary, the first phase shown in FIG. 5A may be termed a learning phase and the models Γ_γtn and Ψ_ψtn at each stage n and each step t may be trained until convergence by solving

$\begin{array}{cc}\underset{{\gamma }_t^n,{\psi }_t^n}{\mathrm{argmin}}{L}_{\mathrm{drift}}^t\left({\prod }_{\mathrm{old}}^{t,0},{\prod }_{\mathrm{old},\left\{f,s\right\}}^{t,n,},{ℱ}_t^0,{ℱ}_t^n;{\gamma }_t^n,{\psi }_t^n\right)& \left(5\right)\end{array}$

where L_drift^t is the loss of the current model on inferred representations of old classes, _t⁰ denotes the set of features extracted using a feature extractor f_Θt at the beginning of the incremental step t (i.e. at optimization step 0), _tⁿ denotes the set of features extracted using f_Θt updated with n>0 optimisation stages from the dataset _t, Π_old^t,0 is the set of prototypes of old classes C_old at the optimization step 0, Π_old^t,n is the set of prototypes updated at the n^th epoch at the step t, γ_tⁿ denotes the parameters of the feature drift model at time step t updated with n>0 optimisation stages and ψ_tⁿ denotes the parameters of the semantic drift model at time step t updated with n>0 optimisation stages. It is noted that _t⁰ and _tⁿ contain only representations of new classes C_t because only _t is available to the step t. Gradient backpropagation may be used for training both models. It is noted that the convergence criterion for a model is early stopping the optimisation of model parameters if the training loss does not change for a fixed/set number of steps.

The second phase shown in FIG. 5B may be termed an inference phase and the models Γ_γtn and Ψ_ψtn which were learnt in the previous phase are fixed and used to estimate the features of the revived evanescent representations F_oldⁿ. Then, the set of old class prototypes Π_old^t,n may be computed by class-wise averaging features sampled from p(F∈F_oldⁿ). Finally, f_θt and h_ϕt are trained on the updated dataset for the time step =_t∪F_oldⁿ for example by a loss function which may be represented as:

$\begin{array}{cc}\underset{{\theta }_t,{\varphi }_t}{\mathrm{argmin}}{L}_{\mathrm{cc}}^t{t}_{};{\theta }_t,{\varphi }_t)& \left(6\right)\end{array}$

FIG. 5C shows a block diagram of an apparatus 100 for training the model as described above and/or implementing the trained model. The apparatus 100 may be any one of: a smartphone, tablet, laptop, computer or computing device, virtual assistant device, a vehicle, a drone, an autonomous vehicle, a robot or robotic device, a robotic assistant, image capture system or device, an augmented reality system or device, a virtual reality system or device, a gaming system, an Internet of Things device, or a smart consumer device (such as a smart fridge). It will be understood that this is a non-exhaustive and non-limiting list of example devices.

The apparatus comprises at least one processor 102 coupled to memory 104. The at least one processor 102 may comprise one or more of: a microprocessor, a microcontroller, and an integrated circuit. The memory 104 may comprise volatile memory, such as random access memory (RAM), for use as temporary memory, and/or non-volatile memory such as Flash, read only memory (ROM), or electrically erasable programmable ROM (EEPROM), for storing data, programs, or instructions, for example.

The apparatus may further comprise at least one image capture device 108 for capturing images or videos to be processed by the ML model 106. The ML model may be an image classification model such as that shown in FIG. 1A which is trained as described in detail above.

The apparatus may further comprise at least one interface 110 for providing a result of the processing by the ML model to a user of the apparatus. For example, the apparatus 100 may comprise a display screen to receive user inputs and to display the results of implementing the ML model 106.

Experimental Results

We evaluate our approach on multiple standard CIL benchmarks, including CIFAR100 [17], TinyImageNet [24] and CUB200-2011 [33] datasets. FIGS. 6A to 8F show the experimental results.

We devise 3 class-incremental setups, where first the framework is trained on half of the available semantic classes (except for one setup on CIFAR 100, where only 40 classes are selected as the first task), and then the remaining class set is evenly divided into respectively 5, 10 or 20 incremental steps (or phases—the terms may be used interchangeably). Class order is selected randomly and then fixed for every class split.

Implementation Details.

ResNet-18 [10] is used as a backbone. The model is trained for 100 epochs (i.e. N=100 and each stage corresponds to one epoch over _t) at each incremental step with Adam optimizer. The learning rate is initially set to 0.001 on CIFAR 100 and TinyImageNet and 0.0001 for CUB datasets. It is decreased by a factor of 0.1 after 45 and 90 epochs, for example as suggested in [43]. Images are cropped to 32×32, 64×64 and 256×256 for CIFAR100, TinyImageNet and CUB200-2011, respectively, and randomly flipped. We apply input and label augmentation, for example as described in [43]. We set batch size to 64, and λ_fkd=10 and λ_rd=10 in all experiments.

We employed lightweight DNNs to identify networks of Γ_γ and Ψ_ψ for modeling feature and representation drifts. In particular, we investigate both the use of GMs with a multi-layer perceptron (MLP) which may have only two layers, and generative models in the form of a condition VAE [30, 42] to implement VMs, whereby hyperparameters are experimentally tuned. In the following sections we will denote a multi-layer perceptron by MLP and a generative model by VAE. The VAE may have encoder and decoder modules composed of a couple of FC layers and the MLP may have two layers.

The fusion loss L_fus^t defined above may be implemented by two methods for an ablation:

• • i) we define ρ by a normalised feature kernel matrix, for example as described in [3] or
  • ii) we define ρ by an identity map, i.e. A=ρ(A).The results obtained by the second method (ii) are marked with a † in the tables shown in the Figures.

Comparisons.

We compare our approach with many methods storing exemplars of old classes (EEIL [4], iCarl [25], UCIR [12]) and other state-of-the-art approaches that instead avoid their use (EWC [16], LwF [19], LwM [8], PASS [43], SDC [40]). All the methods are evaluated with the ResNet18 image classification model and batch size of 64 as described in [15]. We perform gridsearch over key hyperparameters of [2, 1, 9, 3, 6, 7]. As for the exemplar-based methods, we store 20 samples with herd selection, for example as described in [12,25]. In addition, we use the original code of PASS [43]. We evaluate the SDC [40] method by employing the prototype drift compensation proposed in [40] to update prototypes of past classes, and model old-class feature distribution by Gaussians as discussed above. In particular, we employ the original code of [40] to evaluate and compensate for the feature drift of old-class prototypes (i.e. in place of the proposed semantic and feature representation drift models), and we use the estimated up-to-date representations Π_old^t,n by SDC [40] to approximate feature distribution of old classes with a parametrized Gaussian model, i.e. p(F∈_oldⁿ: π_c∈Π_old^t,n)˜N(π_c,σ_c), where N is the normal distribution and σ_c is the standard deviation estimated at step t′. In the following sections we will show how our CIL method outperforms SOTA non-exemplar frameworks, while surpassing also some approaches using exemplars.

Evaluation Metrics.

To evaluate and compare different approaches we resort to the accuracy metric [25], defined as the average top-1 classification accuracy over all classes up to the current incremental step.

TinyImageNet

In FIGS. 6A to 6C, we report per-step incremental accuracy results in graphical form when evaluation is performed on TinyImageNet benchmark. These graphs show the improved performance achieved by our approach with respect to the prior art approaches noted above. These results are also shown in Table 2 in FIG. 6G. For our method, the results are shown for linear and non-linear models as represented by MLP and VAE respectively. For our method, the results are also shown separately for modelling feature drift alone (represented by Feat Drift), modelling semantic drift alone (represented by Sem Drift) and for the fusion model in which feature drift and semantic drift is used (represented by fusion).

We observe that our framework outperforms exemplar-based competitors, while also providing an accuracy boost over state-of-the-art methods not using exemplars [43, 40]. In particular, it is possible to notice that our representation drift modeling yields superior performance with respect to that proposed in [40]. This is especially true when semantic and feature representation shifts are jointly taken into account, showing that they both individually model crucial and complementary information, not fully captured by the framework proposed by [40]. Finally, we remark that employing a probabilistic generative modeling can further boost the final accuracy, especially when multiple incremental steps are performed. Nonetheless, a simple MLP architecture already provides satisfactory results.

CIFAR100

In FIGS. 6D to 6F, we report per-step incremental accuracy results in graphical form when evaluation is performed on the CIFAR100 benchmark. These graphs show the improved performance achieved by our approach with respect to the prior art approaches noted above. These results are also shown in Table 2 in FIG. 6G. Our models (with the best achieved accuracy) outperform the closest SotA (SDC) by 1.75%, 4.6% and 3.07% for 5, 10 and 20 steps. In FIGS. 6D to 6F, for the sake of clearance of presentation of results, we only provide per-step incremental accuracy of Fusion (GM-MLP and VM-VAE). The results show the improved accuracy achieved by our models with respect to the prior art throughout the incremental steps.

CUB200-2011

In FIGS. 7A to 7C, we report final accuracy results in graphical form when evaluation is performed on the CUB200-2011 benchmark. These graphs show the improved performance achieved by our approach with respect to the prior art approaches noted above. These results are also shown in Table 3 in FIG. 7D which uses the same referencing as Table 2 in FIG. 6G.

Here non-exemplar methods provide quite low results, especially when the number of incremental steps is increased. Adopting the method proposed in [40] to compensate for modeling shift of prototypes using a softmax classifier seems to have no beneficial effect, showing that it fails to adequately model semantic drift in a fine-grained classification set-up with high semantic similarity among classes. On the other end, our framework demonstrates to successfully capture model representation drift; by injecting up-to-date knowledge of old classes, in fact, we manage to much more effectively mitigate catastrophic forgetting.

FIGS. 8A to 8F show two different measures of forgetting plotted against number of steps. A base measure of forgetting f_t^c may be computed for each past class c at step t>0 by

$f_t^c=\underset{k<t}{\max }\left(a_k^c-a_t^c\right),\forall c\in {\bigcup }_{k<t}{𝒞}_k$

where a_k^c denotes the top-1 accuracy for class c attained at the step k<t. We then compute the class-wise average of forgetting measures over all past classes at each step. This class-average measure of forgetting (%) is plotted in FIGS. 8A to 8C. The task-wise average at the end of incremental training may also be calculated by computing the averages of forgetting measures at the final step over classes belonging to the same tasks. This per task average forgetting (%) is plotted in FIGS. 8d to 8f.

FIGS. 8A to 8C show how the class-wise average forgetting evolves throughout incremental training when different methods are employed. We observe that the proposed CIL approach based on representation drift modelling mitigates forgetting more efficiently than the other competitors. FIGS. 8D to 8F show how our CIL method causes less overall forgetting that the competing approaches and provides an improved performance equally shared among all tasks. This is especially noticeable when compared with the PASS and SDC methods which induce almost total forgetting of the oldest tasks.

It will be appreciated that the measures of forgetting may also be calculated for the other datasets but for efficiency just the CUB100 results are shown here. Similar improvements are seen with the proposed approach. Only the known method iCARL [9] shows some improvement over the proposed approach for the CIFAR100 and TinyImageNet datasets. However, we note that iCARL [9] yields a lower or comparable classification accuracy when compared to the proposed approach. This suggests that a focus of iCARL [9] on preserving past knowledge (i.e. stability) is accompanied by a less efficient learning of novel classes (i.e. plasticity).

Ablation Study

As explained above, our framework enables implementation of drift models using different GMs and VMs. We studied the accuracy of a GM (MLP) and VM (VAE) for modeling different drifts and their fusion in the results section above. The results suggest that the accuracy of GMs and VMs depends on statistical sufficiency of data which affects capacity of f_θ and learned representations as follows:

• • On the Cifar100 dataset, the GM (MLP) outperforms the VM (VAE) for smaller (e.g. 5) steps, where more classes are observed at each step, compared to the larger (e.g. 20) steps. We conjecture that this result can be attributed to training models using statistically insufficient data representing all classes at each step.
  • On TinyImageNet, containing larger images than Cifar100, the VM (VAE) performs on par with and slightly outperforms the GM (MLP) for smaller steps.
  • On CUB200 containing the largest images, the VM (VAE) outperforms the GM (MLP) for all steps.

The methods described above are elucidated further with an ablation study and related analyses detailed below.

The first ablation study comprises analysing how well semantic drift is modeled and employed to update prototypical class representations. We explore how well representation drifts are modeled and prototypes are updated considering the drifts during incremental learning. For this purpose, we compute low-dimensional embedding of high dimensional representation vectors using the Isomap [31] so that the 2D embeddings of the feature vectors can be visualised. The results are shown in FIGS. 9A to 12B. In the results, we observe that our proposed methods estimate prototypes closer to the target (ground-truth) prototypes compared to the state of the art method PASS [43].

FIGS. 9 and 10 depict the feature representations of samples of the first four learned classes (which for example may be apple, aquarium fish, baby, bear) from the test set (circles) which is CIFAR100. The prototypes which have been computed over the training data are shown as squares and the prototypes which are estimated during training are shown as triangles. Finally, the prototypes computed over the test data are shown as diamonds. FIG. 9 shows the results for the PASS method and FIG. 10 shows the improved results with the method proposed above. In other words, our proposed methods estimate prototypes which are closer to their evanescent versions compared to PASS, following the trajectory of evanescent representations of old classes, without having access to training data of such classes.

FIG. 11 is a feature representation of samples of the first four learned classes from the CIFAR100 test set with the samples shown as circles. The decrease in transparency and increase in brightness indicates that the representations are extracted at progressively increasing incremental steps (i.e. at steps 0, 10 and 20). The prototypes which have been computed over the training data are shown as squares and the prototypes which are estimated during training are shown ad triangles. Finally, the prototypes computed over the test data are shown as diamonds. The prototypes which have been computed over the training data are fixed for the rest of the incremental tasks.

FIGS. 12A and 12B show the same information as FIG. 11 for the test sets TinyImageNet and CUB200 respectively. The prototypes which have been computed over the training data are shown as squares and the prototypes which are estimated during training are shown ad triangles. Finally, the prototypes computed over the test data are shown as diamonds.

FIGS. 9A to 12B show that our proposed method allows the estimation of revived prototypes which are closer to their evanescent versions than when using PASS, especially for the CIFAR100 and CUB200 benchmarks shown in FIGS. 11 and 12B. This shows that we can approximate the trajectory of evanescent representations of old classes, without having access to training data of such categories, by modeling representation drift.

In the next set of analyses, we investigate how our method proposed to model representation drifts, captures and preserves semantic relationships between feature representations of old and new classes. For this purpose, we compute prototypical representations (or prototypes) of novel classes on the training data available at an incremental step t and estimate the revived prototypes of old categories by modeling semantic and feature drifts (employed individually or fused). Then, we express inter-class relationships in the form of Euclidean distances between prototypes of past and new classes, and observe the evolution of such distance throughout an incremental step t. In particular, we compute distance values between pairs of class prototypes at the beginning and at the end of the same incremental step t, and measure their change across the optimisation interval (i.e. in the form of absolute value of the difference between the estimates performed at the beginning and the end of step t.

We compute the Euclidean and cosine distances between estimated (revived) prototypes of old classes (i.e. computed over training data and fixed [43], or updated by [40] or by drift models) and their reference (i.e. evanescent) representations (computed over the test set) at each incremental step. In other words, we measure and analyze representation drift with geometric distance functions. We provide per-step class-wise average distance values (main curves), as well as the maximum and minimum class values that identify, respectively, the upper and lower bound of the shaded regions for each setup. Feature prototypes of old classes are estimated by computing class wise averages of feature representations over training data when they are available at an incremental step, which can then be simply fixed for the rest of the training (as done in PASS [43]), or can be updated by SDC [40] or by the proposed drift models. Evanescent prototypes of the same old classes are instead computed over the test set (unavailable during training).

We replicate the analysis on the CIFAR100 shown in FIGS. 13A and 13B, on the TinyImageNet shown in FIGS. 13C and 13D and on the CUB200 dataset shown in FIGS. 13E and 13F. FIGS. 13A, 13C and 13F plot the average Euclidean distance between estimated prototypes of old classes and their true for each of PASS [43], SDC [40] and the method proposed above. FIGS. 13B, 13D and 13F plot the average cosine similarity between estimated prototypes of old classes and their true representations for each of PASS [43], SDC [40] and the method proposed above. Distances are reported for each incremental step while training modes' on the CIFAR100 dataset with 20 incremental steps.

The results show that our proposed methods can update the prototypes to be closer and better aligned with the reference prototypes compared to the state of the art PASS and SDC methods. Furthermore, we observe that the improved accuracy with respect to the known methods is shared among the three datasets chosen for evaluation. In particular, we notice the remarkable improvement experienced on the CUB200 dataset. Our method, in fact, provides much lower Euclidean distance between revived and evanescent representations, whose average value is kept almost constant as incremental training progresses and new classes are introduced. The same trend can be observed for the cosine similarity of the estimated and evanescent representations, whose value tends to be steady through the incremental training and much closer to the upper bound when our method is adopted. Thereby, our proposed methods reduce the drift between semantic representations in terms of geometric distances (i.e. Euclidean and cosine distances).

In FIGS. 13G to 13L, we compare the normalised Euclidean distance between the estimated and evanescent prototypes for different numbers of total incremental steps. The results using the CIFAR100 dataset are shown in FIGS. 13G and 13H, on the TinyImageNet shown in FIGS. 13I and 13J and on the CUB200 dataset shown in FIGS. 13K and 13L. FIGS. 13G, 13I and 13FK plot the average Euclidean distance between estimated prototypes of old classes and the evanescent prototypes of old classes for each of PASS [43], SDC [40] and the method proposed above using a MLP. FIGS. 13H, 13J and 13L plot the average Euclidean distance between estimated prototypes of old classes and the evanescent prototypes of old classes for each of PASS [43], SDC [40] and the method proposed above using a VAE. As shown, the proposed method always outperforms the known methods. This is particularly noticeable for the 20-step setup on CU200, where our approach jointly shows the largest improvement over the state of the art methods by approximately 20%. In addition, we notice that by employing VM to identify representation drifts, we reach the largest improvement over PASS, which translates into the revived evanescent representations closer to their unknown reference version.

FIGS. 14A to 14F show the change of normalized distance between prototypes in consecutive incremental steps. The results show that our methods enable control of the drift of representations by limiting its change better than the state of the art method PASS. FIGS. 14A and 14D plot the pairwise normalised Euclidean distance between prototypical feature representations of classes introduced in the first incremental phase for the PASS method and our method respectively. FIGS. 14B and 14E plot the pairwise normalised Euclidean distance between prototypical feature representations of classes introduced in the second incremental phase for the PASS method and our method respectively. For each of FIGS. 14A, 14B, 14D and 14E, the Euclidean distance is captured at the beginning (left) and end (mid) of the second phase (CIFAR100, 20 phases). FIGS. 14C and 14F plot the absolute values of the difference for the PASS method and our method respectively.

We focus on analyzing the change of semantic relationship during the first incremental step (i.e. t=1). We notice how prototypes estimated by leveraging the modeled semantic representation drift tend to more effectively preserve inter-class relationships with respect to novel classes. By utilising feature drift alone, in fact, we notice that class representations tend to modify their interconnections. Nonetheless, the proposed model fusion allows to better identify and retain inter-class relationships, whereas keeping prototypes fixed (as in PASS) causes a greater impairment of inter-class relationships as new representations are learned.

In the next set of analyses, we analyse the relationship between semantic drift and model accuracy. Specifically, we analyse the relationship tying the accuracy of the classification model and the normalised distance between the revived and evanescent prototypes of old classes. Normalisation is performed by dividing individual distances computed for single past classes by the average distance among all the past classes. We then provide the average of normalised distance values at each incremental step. We evaluate the accuracy of the proposed method when fusing semantic and feature drift models and adopting GM to identify representation drift, alongside with that of the SDC and PASS.

FIGS. 15A to 15C plot the top-1 accuracy (%) in relation to normalised prototype Euclidean distance to reference (evanescent) representations for each of our method, PASS and SDC methods using the CIFAR100, TinyImageNet and CUB200 datasets respectively. Each point corresponds to a single training phase and the decrease in transparency indicates progressively increasing time steps. For each time step, the reported measures have been averaged over all past classes.

We observe that classification accuracy measured at each incremental step and distance between the estimated and evanescent (also known as ground-truth) prototypes are negatively correlated, with similar trends shared by the different methods being analysed across the CIFAR100, TinyImageNet and CUB200 datasets. It is worth noting that our method and SDC display very similar correlation patterns, whilst the latter reaches lower accuracy and higher distance values at the final incremental step.

The results show that state of the art PASS and SDC models overfit to training data as the models are trained incrementally. For instance, accuracy of PASS and SDC continues to decrease below 50% and their produced prototypical representations diverge from the reference prototype as the incremental steps increase. However, the proposed methods limit the distance between prototypical representations by 0.3 and the accuracy by 50%. Therefore, we argue that our proposed method yields superior performance compared to the state of the art methods of PASS and SDC by more accurately tracking and modeling evanescent old class prototypes.

In the final set of analyses we use statistical analyses of representations. We explore how well the estimated revived prototypes of old classes exemplify feature representations of samples of such categories (which are unavailable during training at incremental steps). To this end, we first compute probability distributions over the set of old classes based on the Euclidean distance between feature representation and class prototypes by:

p _F(c)=exp(−∥F−π_c)∥₂/ζ)/Σ_j exp(−∥F−π_j)∥₂/ζ)

where F∈F_old is the representation of a test sample of _old as extracted by the current feature extractor, {π_j}_j are estimated prototypes of _old and ζ is set to 0.1. It is noted that the expression above is equivalent to the Softmax function and in this experiment we have used ζ set to 0.1 although 0.01 may also be used in different experiments.

We analyse the change of entropy (H) and cross-entropy (CE) of p_F across incremental steps in FIGS. 16A and 16B. The cross-entropy is computed with respect to the one-hotted ground-truth distribution of the labeled input sample corresponding to F∈F_old. In addition, for each old class c∈_old, we compute the mean H and CE of p_F, and then, we average over all F corresponding to the same c. FIG. 16A plots the average entropy of the distribution against incremental step for each of PASS, SDC and our method. FIG. 16B plots the cross-entropy of the distribution against incremental step for each of PASS, SDC and our method. Measures are reported for each incremental step when using the CIFAR100 dataset and 20 incremental steps.

We observe that our method provides higher H and smaller CE compared to PASS, and this trend is shared across CIFAR100, TinyImageNet and CUB200. This result suggests that information capacity of representations learned by our methods increases along with classification accuracy more rapidly compared to the SotA as models are incrementally trained.

To further validate this claim, for each old class c, we compute the probability distribution over feature representations of input samples of all past categories, defined by

${p_c\left(F_i\right)=\exp \left(-F_i-{\pi }_c\right)}_2/\tau )/\sum _j\exp \left(-F_j-{\pi }_c\right){}_2/\tau ),$

where F∈F_old are feature representations of test samples and {π_j}_j are revived old-class prototypes. In addition, τ is set to 0.1. In FIGS. 17A to 17H, we report entropy values of p_c across different incremental steps. FIGS. 17A and 17E plot the class-wise entropy of the feature distribution against incremental step for our method and PASS, respectively. FIGS. 17B and 17F plot the class-wise entropy of the class distribution against incremental step for our method and PASS, respectively. FIGS. 17C and 17G plot the cross-entropy of the class distribution against incremental step for our method and PASS, respectively. FIGS. 17D and 17H plot the class-wise top-1 accuracy against incremental step for our method and PASS, respectively. The measures are reported for each incremental step when using the CIFAR100 dataset and 20 incremental steps.

Once more, we observe that our method causes entropy of p_c to reach higher values compared to PASS, indicating that prototypical representations revived by our method are more informative than their fixed counterparts, while still being representative of the corresponding class, as suggested by the lower CE of p_F.

We also analyse the learning curves of semantic and feature drift models. FIGS. 18A to 18H show the convergence values achieved by the total loss function used to train drift models at different epochs across multiple incremental steps. The Figures are plotted for the CUB100 dataset and a setup employing 5 incremental steps. FIGS. 18A and 18B plot the semantic drift loss for the first task and the fifth task when using the MLP adaptation described above. FIGS. 18C and 18D plot the semantic drift loss for the first task and the fifth task when using the VAE adaptation described above. FIGS. 18E and 18F plot the feature drift loss for the first task and the fifth task when using the MLP adaptation described above. FIGS. 18G and 18H plot the feature drift for the first task and the fifth task when using the VAE adaptation described above.

For semantic drift models, we show the results for the total loss (denoted by Fusion (VAE/MLP)) comprising L_s and L_fus. As for feature drift models, we present the individual behaviour of the loss L_f denoted by Only Feat Drift (VAE/MLP) (without aggregating the loss L_fus), along with combination of the loss functions L_f and L_fus denoted by Fusion (VAE/MLP).

In the results, we observe that the convergence point stabilises in all the analysed setups as learning progresses during each incremental step. This suggests that, as the classification model learns representations of new classes throughout an incremental step, drift models converge to a stable configuration, in which they can provide a robust estimate of representation drift. Moreover, we observe that the VMs (VAE) converge faster than GMs (MLP). In the analyses, we first observe that our methods provide distributions with higher entropy and smaller cross-entropy compared to the state of the art PASS and SDC methods. This result suggests that information capacity of representations learned by our methods increases more as the models are incrementally learned. In addition, our methods provide better semantic representations which represent categories with higher confidence (measured by class-wise cross-entropy) compared to PASS and SDC.

Those skilled in the art will appreciate that while the foregoing has described what is considered to be the best mode and where appropriate other modes of performing present techniques, the present techniques should not be limited to the specific configurations and methods disclosed in this description of the preferred embodiment. Those skilled in the art will recognise that present techniques have a broad range of applications, and that the embodiments may take a wide range of modifications without departing from any inventive concept as defined in the appended claims.

Contributions of the approach proposed above to overcome all the aforementioned limitations and address the CIL problem can be summarized as follows:

• • We provide a probabilistic approach for CIL as described above, rigorously highlighting the source of catastrophic forgetting as the unavailability of an up-to-date estimate of distribution p(F_old) of feature representations gF_old learned using old classes C_old.
  • We model the change of representations of classes as semantic drift (i.e., due to the change of semantic categories learned at different incremental steps) and propose an effective implementation to estimate it.
  • We model the change of feature representations as feature drift (i.e, due to constantly evolving feature representations of different patterns learned from data by the model while training models) and propose an effective implementation to estimate it.
  • We devise a joint framework where semantic and feature drifts are fused. The resulting combined representation drift is used to obtain an up-to-date estimate of feature distributions for C_old to be exploited to preserve C_old knowledge when learning new representations for C_t, ∀t.In summary, the proposed method addresses catastrophic forgetting, through continually learning using only incrementally available datasets. The method also avoids data storing issues, by re-estimating old-class data distributions. As demonstrated by the results, the proposed method achieves high performance, by leveraging updated representations of past data. There is also enhanced semantic understanding, by modeling inter-class relationships and evolution of latent representations.A full list of the references mentioned in the description is provided in the list below. The teaching of these references is incorporated herein by reference:
• [1] Hongjoon Ahn, Jihwan Kwak, Su Fang Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In ICCV, 2021.
• [2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elho-seiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, 2018.
• [3] Anonymous. Feature kernel distillation. In Submitted to The Tenth International Conference on Learning Representations, 2022. under review
• [4] Francisco M. Castro, Manuel J. Marin-Jimenez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), 2018.
• [5] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l:Contrastive continual learning. In ICCV, 2021.
• [6] Arslan Chaudhry, Puneet Kumar Dokania, Thalaiyasingam Ajanthan, and Philip H. S. Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In ECCV, 2018.
• [7] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 2021.
• [8] Prithviraj Dhar, Raj at Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.
• [9] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), 2020.
• [10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
• [11] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2015.
• [12] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.
• [13] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class-incremental learning. In CVPR, 2021.
• [14] Metod Jazbec, Matt Ashman, Vincent Fortuin, Michael Pearce, Stephan Mandt, and Gunnar Ra'tsch. Scalable gaussian process variational autoencoders. In AISTATS, 2021.
• [15] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. In ICLR, 2018.
• [16] James Kirkpatrick, Razvan Pascanu, Neil C. Rabinowitz, Joel Veness,v Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. CoRR, abs/1612.00796, 2016.
• [17] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009.
• [18] Matthias De Lange and Tinne Tuytelaars. Continual prototype evolution: Learning online from non-stationary data streams. In ICCV, 2021.
• [19] Zhizhong Li and Derek Hoiem. Learning without forgetting. In Proceedings of the European Conference on Computer Vision (ECCV), 2016.
• [20] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In CVPR, 2021.
• [21] James Lucas, George Tucker, Roger B Grosse, and Mohammad Norouzi. Don't blame the elbo! a linear vae perspective on posterior collapse. In NeurIPS, 2019.
• [22] Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey. Neurocomputing, 469:28-51, 2022.
• [23] German I. Parisi, Ronald Kemker, Jose L. Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54-71, 2019.
• [24] Hadi Pouransari and Saman Ghili. Tiny imagenet visual recognition challenge. In CS231N course, Stanford Univ., Stanford, Calif., USA, 2014.
• [25] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
• [26] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In NeurIPS, 2017.
• [27] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. On learning the geodesic path for incremental learning. In CVPR, 2021.
• [28] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In ICCV, 2021.
• [29] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, editors, NeurIPS,2017
• [30] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems (NeurIPS), 2015.
• [31] Joshua B. Tenenbaum, Vin de Silva, and John C. Langford. A global geometric framework for nonlinear dimensionality reduction. Science, 290(5500):2319-2323, 2000.
• [32] Eli Verwimp, Matthias De Lange, and Tinne Tuytelaars. Rehearsal revealed: The limits and merits of re-visiting samples in continual learning. In ICCV, 2021
• [33] Catherine Wah, Peter Welinder Steve Branso and, Pietro Perona, and Serge Belongie. The CaltechUCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011.
• [34] Chenshen Wu, Luis Herranz, Xialei Liu, Yaxing Wang, Joost van de Weijer, and Bogdan Raducanu. Memory replay gans: Learning to generate new cat-egories without forgetting. In NeurIPS, 2018
• [35] Guile Wu, Shaogang Gong, and Pan Lid. Striking a balance between stability and plasticity for class-incremental learning. In Proceedings of the International Conference on Computer Vision (ICCV), 2021.
• [36] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019.
• [37] Ye Xiang, Ying Fu, Pan Ji, and Hua Huang. Incremental learning using conditional adversarial networks. In ICCV, 2019.
• [38] Yaniv Yacoby, Weiwei Pan, and Finale Doshi-Velez. Failure modes of variational autoencoders and their effects on downstream tasks, 2021
• [39] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: dynamically expandable representation for class incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021.
• [40] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020.
• [41] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In Doina Precup and Yee Whye Teh, editors, ICML, 2017
• [42] Shengjia Zhao, Jiaming Song, and Stefano Ermon. Infovae: Information maximizing variational autoencoders. CoRR, abs/1706.02262, 2017.
• [43] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and Cheng-Lin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021.

Claims

1. A computer-implemented method for training a classifier using class incremental learning, CIL, wherein the classification model is a machine learning, ML, model comprising a feature extraction model and a classification model and wherein at each incremental time step in the CIL a new training dataset having a set of samples with class labels is obtained and an optimisation over a plurality of stages is applied, the training method comprising for each incremental step: extracting, using the feature extraction model, a first set of features representing the samples within the new training set at the beginning of the incremental step and a second set of features representing the samples within the training set at a subsequent stage within the optimisation;obtaining a first set of old class prototypes for the beginning of the incremental step, wherein each class prototype is a semantic representation of the sets of features of the samples which have the same class label and wherein each old class prototype represents old samples which are samples that are not present in the new training set;learning a feature drift model using the extracted first and second sets of features, wherein the feature drift model estimates the drift between the features representing each sample as the classification model is updated;learning a semantic drift model using the set of old class prototypes and first set of features, wherein the semantic drift model estimates the drift between the class prototypes for each class as the classification model is updated;inferring at least one of a second set of old class prototypes and a set of old features using the learnt feature drift model and the learnt semantic drift model, wherein the second set of prototypes are semantic representations of revived old samples at a subsequent stage within the optimisation and the set of old features are the feature representations of revived old samples at the subsequent stage within the optimisation; andusing at least one of the inferred second set of old class prototypes and set of old features to update the ML model; andwhen the final incremental step is completed, outputting the trained classification model.

2. The method of claim 1, wherein the classifier is an image classifier and the dataset comprises a plurality of samples in the form of images.

3. The method of claim 1, wherein the feature drift model uses a linear deep neural network and learning the feature drift model comprises minimising the error between the extracted second set of features and a third set of features which are the estimate of these features using the feature drift model.

4. The method of claim 1, wherein the feature drift model uses a variational model and learning the feature drift model comprises updating the weights γtn of the model to maximize the likelihood that a set of features representing a sample is in the second set of features when it is in the first set of features, namely: maximise p(F∈tn|F∈t0;γtn)where t0 denotes the extracted first set of features, tn denotes the extracted second set of features and F is the up-to-date feature representations from both the set of features of new classes and old classes.

5. The method of claim 1, wherein the semantic drift model uses a linear deep neural network and learning the semantic drift model comprises minimising the error between the first set of old class prototypes and a third set of old class prototypes which are the estimate of these old class prototypes using the semantic drift model.

6. The method of claim 1, wherein the semantic drift model uses a variational model and learning the semantic drift model comprises updating the weights ψ of the model to maximize the likelihood that a class prototype representing a sample is in the first set of old class prototypes when the features of the sample are in the first set of features namely: maximise p(π∈Πoldt,n=0|F∈t0;ψtn)where t0 denotes the first set of features, Πoldt,0 is the first set of old class prototypes, F is the up-to-date feature representation and π is an up-to-date class prototype.

7. The method of claim 3, wherein the inferring step comprises inferring the second set of old class prototypes using the feature drift model to map the first set of old class prototypes to the second set of old class prototypes.

8. The method of claim 4, wherein the inferring step comprises inferring the set of old features using the trained feature drift model to provide an approximation for the probability distribution of p(F∈oldn|F∈old0) where old0 is a first set of features for the old samples at the beginning of the incremental step and oldn is a second set of features for the old samples at the subsequent stage of optimisation.

9. The method of claim 5, wherein the inferring step comprises inferring the second set of old class prototypes using the semantic drift model to map the extracted second set of features to the second set of old class prototypes.

10. The method of claim 6, wherein the inferring step comprises inferring the set of old features using the trained semantic drift model to compute the posterior probability p(π∈Πoldt,n>0|F∈tn; ψ), where Πoldt,n is the second set of old class prototypes, tn denotes the extracted second set of features and ψ are the weights of the semantic drift model.

11. The method of claim 1, wherein using the inferred second set of old class prototypes to update the ML model comprises training the feature extraction model and the classification model based on at least one of the inferred second set of old class prototypes and the inferred set of old features.

12. The method of claim 11, wherein training the feature extraction model comprises minimising a distillation loss which is a distance between a first set of representations extracted from the current dataset using the current feature extraction model and a second set of representations extracted from the current dataset using the feature extraction model from the previous incremental time step.

13. The method of claim 11, wherein training the classification model comprises using at least one of a cross-entropy loss and a representation drift loss function

14. The method of claim 1, wherein the extracting, learning, obtaining and using steps are repeated for each optimisation stage.

15. A method of classifying a sample, the method comprising inputting a sample;extracting a plurality of features from the sample using a feature extraction model of a classifier which has been trained according to the method of any one of the preceding claims;classifying the sample based on the extracted plurality of features using a classification model of a classifier which has been trained according to the method of any one of the preceding claims; andoutputting the classification of the sample.

16. A non-transitory data carrier carrying processor control code to implement the methods as defined in claim 1.