Patent Application
20230385384
The present invention relates to Machine Learning.
Machine learning is used as a means for classifying data in many fields. One very common example is in the classification of images. Images are generally characterised by carrying rich patterns, which contain a significant portion of features that do not necessarily dependent on the context, such as for example the place of acquisition, time and various environmental conditions.
Other types of data are sparse, and their patterns may have a particular meaning only in the specific context from which the data originate. For example, we may consider data obtained from tracking systems using radars, cameras and the like, which typically comprise sequences of locations, speeds and orientations.
a,
1
b, 1c and 1d present scenarios illustrating this point in relation to naval tracking data.
As shown in
As shown in
Accordingly, in
As shown in
As shown in
As such, once again, in
This dependence on Context and sparse data of the kind presented above means that certain common machine learning approaches, such as for example Neural Networks might not be best suited.
Probabilistic Graphical Models meanwhile may be seen as better suited to such fields due to their ability to efficiently model the context and causal relations. They facilitate inclusion of expert knowledge and can automatically learn the specific properties of a context, however the size and complexity of Probabilistic Graphical Models grows with the number of relations and the states of variables, making the learning of such models challenging to efficiently capture intricate behaviours requiring higher modelling resolution, such as U-turns, ZIG-ZAGs and the like in the context presented above.
Attempts to combine different types of models through machine learning are known for example from the article by Y. Bengio, R. De Mori, G. Flammia, and R. Kompe entitled “Global optimization of a neural network-hidden Markov model hybrid.” Published in IEEE Transactions on Neural Networks, 3(2): 252-259, 1992 and Diederik P Kingma, Danilo J Rezende, Shakir Mohamed and Max Welling: Semi-Supervised Learning with Deep Generative Models, NIPS, 2014.
It is accordingly desired to develop new Machine learning structures better addressing the foregoing considerations.
In accordance with the present invention in a first aspect there is provided a method of building a computer implemented data classifier for classifying data from a specified context (C1), the method comprising the steps of:
In a development of the first aspect, one or more Observable variables (Var1, Var2, Var3, . . . , VarN) of a Probabilistic Graphical Model are directly dependent on the class Variable, and one or more Latent variables.
In a development of the first aspect, the Probabilistic Graphical Model is extended with one or more Extension variables (VarX1, VarX2 . . . , VarXN), whereby the Extension variables are directly dependent on the class variable, one or more Latent variables and possibly one or more Observable variables.
In a development of the first aspect, the step of obtaining a Probabilistic Graphical Model comprises training the Probabilistic Graphical Model with the first training data (D1.1) from the specified context (C1), the first training data comprising data corresponding to a first set of one or more Observable variables (Var1, Var2, Var3, . . . , VarN), whereby embedding training using the embedding training set modifies only the parameters corresponding to the dependencies between the Extension variables and other variables in the extended Probabilistic Graphical Model.
In a development of the first aspect, embedding training using the embedding training set modifies all parameters corresponding to the dependencies between all variables in the extended Probabilistic Graphical Model.
In a development of the first aspect, there are provided one or more further machine learning models each said further machine learning model comprising said second set of Observable variables (VarA, VarB, . . . , VarZ), and each said further machine learning model output O1.2 comprising probabilities corresponding to the values of said Extension variables (VarX1, VarX2, . . . , VarXN) of said extended Probabilistic Graphical Model, and wherein
In a development of the first aspect, there are provided one or more further machine learning models each said further machine learning model comprising said second set of Observable variables (VarA, VarB, . . . , VarZ), and each said further machine learning model output O1.2 comprising values that are not probabilities, the values corresponding to the states of Observed Extension Variables, a subset of the Extension variables (VarX1, VarX2 . . . , VarXN), whereas the rest of the Extension variables are Latent Extension Variables, wherein the Observed Extension Variables are conditioned on the Latent Extension Variables, and the step of performing an embedding training of a Probabilistic Graphical Model is performed such that for each Observed Extension Variable and each Latent Extension Variable a specific probability table is obtained.
In a development of the first aspect, the Extension variables are directly dependent on the class variable
In a development of the first aspect, the second training data (D2) belongs to the specified context (C1).
In a development of the first aspect, the step of training the machine learning model comprises incorporating the machine learning model as the Latent representation of an autoencoder.
In a development of the first aspect, the machine learning model is trained in an unsupervised mode.
In a development of the first aspect, the context comprises the conditions corresponding to the geographic locations, time and the type of moving entities in a physical space under which the data is sampled.
In a development of the first aspect, the first training data, second training data and third training data comprise kinematic data for moving entities in a physical space.
In a development of the first aspect, the first training data, second training data and third training data further comprise images, video streams, sound or electromagnetic signatures.
In accordance with the present invention in a second aspect there is provided method of classifying data comprising presenting the data to a classifier in accordance with the first aspect.
In accordance with development of the method of the first or second aspects, the method is applied to classification of targets in combat management systems, or in processing of sensor observations or detections of moving targets, wherein the first set of Observable variables (Var1, Var2, Var3, . . . , VarN) and the second set of Observable variables (VarA, VarB, VarC, . . . , VarZ) correspond to (i) the outputs of various sensors perceiving the targets and (ii) outputs of sources describing the environmental conditions and wherein the dependencies between the Observable variables, the class variable and each Extension variable describe the correlations between the context, the observations and the target class, enabling classification of a target, prediction of its states or detection of anomalous target states.
In accordance with the present invention in a third aspect the method of the first or second aspect is applied to detection of anomalies in IT systems, cyber physical systems and detection of cyber attacks, wherein the first set of Observable variables (Var1, Var2, Var3, . . . , VarN) and the second set of Observable variables (VarA, VarB, VarC, . . . , VarZ) correspond to the readings from various IDS probes at different system levels and wherein the dependencies between the Observable variables, the class variable and each Extension variable describe the correlations between different components of the overall system, such that the states of unobservable components can be predicted or anomalous states of components can be detected.
In accordance with the present invention in a fourth aspect there is provided a data processing system comprising means for carrying out the steps of the method of any of the first, second, or third aspects.
In accordance with the present invention in a fifth aspect there is provided a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any of the first, second, third or fourth aspects.
In accordance with the present invention in a seventh aspect there is provided a computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of any of the first, second, third or fourth aspects.
The invention will be better understood and its various features and advantages will emerge from the following description of a number of exemplary embodiments provided for illustration purposes only and its appended Figures in which:
In general terms, it is desired to implement a transfer learning mechanism, whereby detailed behaviours learned by a Neural Network or the like may be reused in a different context, whose characteristics are captured in a Probabilistic Graphical Model. Transfer learning mechanisms are conventionally used in pure Deep Neural Networks, where parts of one Neural Network are transferred to a different Neural Network. Incorporating Neural Network elements into a Probabilistic Graphical Model to achieve a hybrid model requires different approaches.
In contrast to prior art methods, embodiment of the present invention make use of arbitrarily complex PGM and introduces special patterns with Latent variables enabling efficient embedding of machine learned components and automated learning of the context. Moreover, embodiments support simultaneous or gradual integration of multiple, very different types of machine learning components, which can be also carried out in a fully unsupervised fashion.
In Bayesian Networks, an important class of Probabilistic Graphical Models used for the illustration, Graphs encode the types of dependencies between the variables (qualitative domain knowledge), while Conditional probability tables encode the strength of dependencies. Graphs are often transferable, being the same for all contexts, while the Conditional probability tables are NOT transferable and must be relearned for each context.
A neural network meanwhile may support efficient training of Fine grained/high resolution models, such as those of behaviours (U-turns, ZIG-ZAG, . . . ). The training can be based on supervised OR unsupervised learning. This learning may be valid under different conditions and consequently can be reused in different contexts, however unsupervised learning results in models capturing “Tacit” knowledge—that is not necessarily comprehensible a posteriori.
In accordance with embodiments, the objective of merging neural network learning with a Probabilistic Graphical Model may be achieved by using a special modelling pattern/harness in the Probabilistic Graphical Model supporting automated learning of relations between embedded features, the classes and the context.
In particular,
The context may comprise the conditions corresponding to the geographic locations, time and the type of moving entities in a physical space under which the data is sampled.
In accordance with a first step as illustrated in
A Special modelling pattern enables automated context learning by means of a graph structure in the Probabilistic Graphical Model, where observed variables directly depend on the class variable and one or more Latent variables and the Latent variables may in some cases directly influence the class variable or vice versa. The Latent variables represent context. In
It will be appreciated that while in some embodiments obtaining the Probabilistic Graphical Model 210 may involve actually training a Probabilistic Graphical Model from the data D1.1, the Probabilistic Graphical Model may comprise a predefined “off the shelf” Probabilistic Graphical Model for a particular context, or may be defined manually by directly defining the variables and manually setting the respective probability tables.
Where the Probabilistic Graphical Model is trained for the purposes of an embodiment, this training may comprise the application of an Expectation-maximization algorithm, a gradient descent optimization method, or other training technique as known in the art.
Probabilistic Graphical Model may be of any type as may occur to the skilled person. In particular, the Probabilistic Graphical Model may comprise a Bayesian network, whereby the parameters comprise prior probabilities and conditional probabilities for each variable.
In accordance with a second step as illustrated in
The machine learning model may comprise any Machine learning model as will be apparent to the skilled person, such as a Decision Tree structure, Hidden Markov Model, Support Vector Machine, or a further Probabilistic Graphical Model, or a neural network.
Optionally, the machine learning model may be trained in an unsupervised mode. For example, the training of the machine learning model may comprise an Autoencoder or a Variational Autoencoder comprising a set of Latent variables corresponding to the machine learning model outputs.
It will be appreciated that while in some embodiments obtaining the machine learning model 220 may involve actually training a machine learning model from the data D2, the machine learning model 220 may comprise a predefined “off the shelf” machine learning model for a particular context, or may be defined manually by directly defining the variables and manually setting the respective probability weightings with regard to the other variables.
In some variants, the second training data (D2) may belong to the first context C1. In other variants, the second training data (D2) may belong to a further context (C2) different from the specified context (C1).
In accordance with a third step as illustrated in
Where the machine learning model does not output probabilities, for example in case the output corresponds to the “Latent space representation” as produced by an autoencoder, the Extension Variables are arranged differently.
Like numbered features correspond generally to those presented with respect to the previous Figures.
In
The skilled person will appreciate that a given implementation may comprise any or all of these interface types in any combination.
Moreover, the Probabilistic Graphical Model may be extended with one or more Extension variables (VarX1, VarX2, VarX3 . . . , VarXN), whereby the Extension variables are directly dependent on the class variable, one or more Latent variables and possibly one or more Observable variables. Specifically, as shown in
Naturally any other configuration may be envisaged as dictated by the structure of the elements used and the characteristics of the underlying data.
In the variant as shown in
Accordingly, Extension variables may be directly dependent on the class variable. In some embodiments, this may be the case for all Extension variables, or all observed Extension variables. Where embodiments define Latent Extension variables, the observed Extension variable may be dependent on the Latent Extension variables only. Still further, all observed Extension variables may be dependent on the class variables and additionally, in some or all cases, on one or more other Latent Extension variables.
The skilled person will appreciate that structures may combine the approaches of
In accordance with a fourth step as illustrated in
The step of training or embedding training the Probabilistic Graphical Model may comprise the application of an Expectation-maximization algorithm.
The step of training or embedding training the Probabilistic Graphical Model may comprise the application of gradient decent optimization method.
The parameters may comprise priors and conditional probabilities for each variable.
Where the context comprises the conditions corresponding to the geographic locations, time and the type of moving entities in a physical space under which the data is sampled, and the first training data, second training data and third training data may comprise kinematic data for moving entities in a physical space.
The first training data, second training data and third training data further comprise images, video streams, sound or electromagnetic signatures.
The step of obtaining an extended Probabilistic Graphical Model may comprise applying an embedding training to a predefined Probabilistic Graphical Model with training data sampled from the specified context (C1) comprising data corresponding to one or more Observable variables (Var1, Var2, Var3, . . . , VarN), in which case the embedding training using the embedding training set may modify only the parameters corresponding to the dependencies between the Extension variables and other variables in the extended Probabilistic Graphical Model. A practical benefit of such an approach is that the embedding training can be carried out in a fully unsupervised way, without using any class labels corresponding to the data patterns in the first training data (D1.1).
Alternatively, the embedding training may modify all parameters corresponding to the dependencies between all variables in the extended Probabilistic Graphical Model. In such a case the data patterns in the first training data (D1.1) must be associated with the class labels.
While
As shown in
Similarly, each further machine learning model output O1.2 may comprise values that are not probabilities corresponding to the states of Observed Extension Variables, a subset of the Extension variables (VarVarX1, VarX2 . . . , VarXN), whereas the rest of the Extension variables are Latent Extension Variables, wherein the Observed Extension Variables are conditioned on the Latent Extension Variables. As such, the step of performing an embedding training of a Probabilistic Graphical Model is performed such that for each Observed Extension Variable and each Latent Extension Variable a specific probability table is obtained.
As shown, the method begins at step 300 before proceeding to step 305 at which a Probabilistic Graphical Model comprising a set of variables comprising a first set of Observable variables (Var1, Var2, Var3, . . . , VarN), and a class variable is obtained, whereby the probabilistic model comprises parameters defining dependencies between the variables of the set of variables.
The method next proceeds to a step 310 of obtaining a machine learning model that is trained on second training data (D2) corresponding to a second set of Observable variables (VarA, VarB, . . . , VarZ).
The method next proceeds to a step 315 of extending the Probabilistic Graphical Model to comprise one or more Extension variables (VarVarX1, VarX2, . . . , VarXN), where some or all of the Extension variables correspond to the outputs of the machine learning model.
The method next proceeds to a step 320 of performing an embedding training of the extended Probabilistic Graphical Model on the basis of an embedding training set of data, the embedding training set comprising first training data (D1.1) of data from the specified context (C1) and an inferred machine learning model output (O1.2) inferred by the machine learning model from third training data (D1.2) from context C1 comprising the second set of Observable variables (VarA, VarB, . . . , VarZ), whereby third training data set D1.2 is sampled from the context C1 (or another context C2 as discussed herein) together with the first training data D1.1, to obtain an enhanced Probabilistic Graphical Model comprising parameters defining dependencies between the Observable variables, the class variable and each Extension variable.
It will be appreciated that this approach provided important benefits. The combination of Machine learning techniques such as Neural Network based techniques with Probabilistic Graphical Models may allow unsupervised learning for the fusion of the feature. During the learning process part of the data is injected into the Probabilistic Graphical Model directly while the other part is “compressed” through the feature embedding/classification component prior to injecting into the Probabilistic Graphical Model. This is because the Expectation—maximization algorithm can carry out general inference about any unobserved variable during learning.
New features may be added without re-learning the entire model, so that it becomes possible for example to easily add new features corresponding to new data sources, such as sensors, as they become available.
The resulting classifier can work with incomplete data (e.g. if a feature is disabled), without any data imputation, hence a robust solution offering graceful degradation is provided.
The described approach is suitable for a generic Probabilistic Graphical Model, imposes no pre-constraints on the type of the inputs, and allows for independent optimization of components. Moreover, different machine learned models can be added to the overall solution over time, as they become available, without the need to retrain the entire set of previously known models and a significant portion of the Probabilistic Graphical Model's parameters.
A classifier obtained as described herein may be used to classify data by presenting data thereto.
Embodiments have been described above in terms of Multi-Loop Bayesian Networks, which offer advantageous characteristics in certain contexts. The presented concepts can be extended to arbitrary classes of Probabilistic Graphical Models, such as any type of Bayesian Network, including for example Dynamic Bayesian Networks and Markov Networks.
Applications of embodiments have been mentioned in the context of the processing of geographical information. It will be appreciated that there exist countless other contexts in which the mechanisms described herein may be particularly useful. Another example may concern the detection of anomalies in IT systems, cyber physical systems or the detection of cyber attacks. In such a context, the first set of Observable variables (Var1, Var2, Var3, . . . , VarN) and the second set of Observable variables (VarA, VarB, VarC, . . . , VarZ) may correspond to the readings from various IDS (Intrusion Detection System) probes at different system levels. The dependencies between the Observable variables, class variable and each Extension variable may then describe the correlations between different components of the overall system, such that the states of unobservable components can be predicted or anomalous states of components can be detected.
This may comprise the further step of displaying information on anomalies and/or cyber attacks on a display, wherein preferably the detected anomalies and cyber attacks are labelled or it is otherwise indicated which type of anomalies or cyber attacks are detected.
A still further application may comprise classification of targets in combat management systems, or in processing of sensor observations or detections of moving targets. In such a context, the first set of Observable variables (Var1, Var2, Var3, . . . , VarN) and the second set of Observable variables (VarA, VarB, VarC, . . . , VarZ) may correspond to (i) the outputs of various sensors perceiving the targets and (ii) outputs of sources describing the environmental conditions. The dependencies between the Observable variables, the class variable and each Extension variable may then describe the correlations between the context, the observations and the target class, enabling classification of a target, prediction of its states or detection of anomalous target states. As such, embodiments may comprise a system such as a radar processing system, combat management system or sensor processing system, comprising a processor or other components adapted for to implement the mechanisms described herein. In particular, there may be provided a vehicle such as a ship, for example a war ship, comprising such a system.
This may comprise the further step of displaying the targets and/or the target states on a display, wherein preferably the targets are labelled or it is otherwise indicated which type of targets are displayed.
The disclosed methods can take form of an entirely hardware embodiment (e.g. FPGA), an entirely software embodiment or an embodiment containing both hardware and software elements. Software embodiments include but are not limited to firmware, resident software, microcode, etc. The invention can take the form of a computer program product accessible from a computer-usable or computer-readable medium providing program code for use by or in connection with a computer or an instruction execution system. A computer-usable or computer-readable can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The medium can be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system (or apparatus or device) or a propagation medium.
Accordingly, there is provided a data processing system comprising means for carrying out the steps of the method as described above, for example with reference to
The data processing system may comprise a display and/or display interface for displaying results of determinations made in accordance with embodiments for example as described above, for example displaying combat management systems targets and/or the target states, wherein preferably the targets are labelled or it is otherwise indicated which type of targets are displayed. Similarly, such a display and/or display interface may be adapted for displaying information on anomalies and/or cyber attacks, wherein preferably the detected anomalies and cyber attacks are labelled or it is otherwise indicated which type of anomalies or cyber attacks are detected.
Similarly, there is provided a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method as described above, for example with reference to
Similarly, there is provided a computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method as described above, for example with reference to
In particular, a method of building a computer implemented data classifier for classifying data from a certain context is provided, whereby the classifier is based on a model obtained by transfer learning combining Probabilistic Graphical Models (PGM) and arbitrary, context independent machine learned models enabled by special modelling patterns, where variables representing outputs of machine learned models are added to the PGM.
These methods and processes may be implemented by means of computer-application programs or services, an application-programming interface (API), a library, and/or other computer-program product, or any combination of such entities.
It will be understood that the configurations and/or approaches described herein are exemplary in nature, and that these specific embodiments or examples are not to be considered in a limiting sense, because numerous variations are possible. The specific routines or methods described herein may represent one or more of any number of processing strategies. As such, various acts illustrated and/or described may be performed in the sequence illustrated and/or described, in other sequences, in parallel, or omitted. Likewise, the order of the above-described processes may be changed.
The subject matter of the present disclosure includes all novel and non-obvious combinations and sub-combinations of the various processes, systems and configurations, and other features, functions, acts, and/or properties disclosed herein, as well as any and all equivalents thereof.
