RAPID IMAGE ANNOTATION VIA BRAIN STATE DECODING AND VISUAL PATTERN MINING

Abstract

Human visual perception is able to recognize a wide range of targets but has limited throughput. Machine vision can process images at a high speed but suffers from inadequate recognition accuracy of general target classes. Systems and methods are provided that combine the strengths of both systems and improve upon existing multimedia processing systems and methods to provide enhanced multimedia labeling, categorization, searching, and navigation.

Description

BACKGROUND

As the volume of digital multimedia collections grow, techniques for efficient and accurate labeling, searching and retrieval of data from those collections have become increasingly important. As a result, tools such as multimedia labeling and classification systems and methods that allow users to accurately and efficiently categorize and sort such data have also become increasingly important. Unfortunately, previous labeling and classification methods and systems tend to suffer deficiencies in several respects, as they can be inaccurate, inefficient and/or incomplete, and are, accordingly, not sufficiently effective to address the issues associated with voluminous collections of multimedia.

Various methods have been used to improve the labeling of multimedia data. For example, there has been work exploring the use of user feedback to improve the image retrieval experience. In some systems, relevance feedback provided by the user is used to indicate which images in the returned results are relevant or irrelevant to the users' search target. Such feedback can be indicated explicitly (by marking labels of relevance or irrelevance) or implicitly (by tracking specific images viewed by the user). Given such feedback information, the initial query can be modified. Alternatively, the underlying features and distance metrics used in representing and matching images can be refined using the relevance feedback information. Ultimately, though, the manual labeling by humans of multimedia data, such as images and video, can be time consuming and inefficient, particularly when applied to large data libraries. Some solutions to the problems described above are disclosed in PCT Patent Application No. PCT/US09/069,237, filed on Dec. 22, 2009, the entirety of which is incorporated herein by reference.

The human brain is an exceptionally powerful visual information processing system. Humans can recognize objects at a glance, under varying poses, illuminations and scales, and are able to rapidly learn and recognize new configurations of objects and exploit relevant context even in highly cluttered scenes. While human visual systems can recognize a wide range of targets under challenging conditions, they generally have limited throughput. Human visual information processing happens with neurons which are extremely slow relative to state-of-the-art digital electronics—i.e. the frequency of a neuron's firing is measured in Hertz whereas modern digital computers have transistors which switch at Gigahertz speeds. Though there is some debate on whether the fundamental processing unit in the nervous system is the neuron or whether ensembles of neurons constitute the fundamental unit of processing, it is nonetheless widely believed that the human visual system is bestowed with its robust and general purpose processing capabilities not from the speed of its individual processing elements but from its massively parallel architecture.

Computer vision systems present their own unique benefits and potential issues. While computer vision systems can process images at a high speed, they often suffer from inadequate recognition accuracy for general target classes. Since the early 1960's there have been substantial efforts directed at creating computer vision systems which possess the same information processing capabilities as the human visual system. These efforts have yielded some successes, though mostly for highly constrained problems. One of the challenges in prior research has been in developing a machine capable of general purpose vision and mimicking human vision. Specifically, an important property of the human visual system is its ability to learn and exploit invariances.

SUMMARY

Both human and computer vision systems offer their own unique benefits and disadvantages. The presently disclosed subject matter combines the benefits of brain state decoding and visual content analysis to improve multimedia data processing efficiency.

Certain embodiments of the disclosed subject matter use brain signals measured by electroecephalopgraphy (“EEG”) to detect and classify generic objects of interest in multimedia data.

Certain embodiments of the disclosed subject matter are designed to facilitate rapid retrieval and exploration of image and video collections. The disclosed subject matter incorporates graph-based label propagation methods and intuitive graphic user interfaces (“GUIs”) that allow users to quickly browse and annotate a small set of multimedia data, and then in real or near-real time provide refined labels for all remaining unlabeled data in the collection. Using such refined labels, additional positive results matching a user's interest can be identified. Such a system can be used, e.g., as a bootstrapping system for developing additional target recognition tools needed in critical image application domains such as in intelligence, surveillance, consumer applications, biomedical applications, and in Internet applications.

Starting with a small number of labels, certain disclosed systems and methods can be implemented to propagate the initial labels to the remaining data and predict the most likely labels (or scores) for each data point on the graph. The propagation process is optimized with respect to several criteria. For example, the system can be implemented to consider factors such as: how well the predictions fit the already-known labels; the regularity of the predictions over data in the graph; the balance of labels from different classes; if the results are sensitive to quality of the initial labels and specific ways the labeled data are selected.

The processes providing the initial labels to label propagation systems can come from various sources, such as other classifiers using different modalities (for example, text, visual, or metadata), models (for example, supervised computer vision models or a brain computer interface), rank information regarding the data from other search engines, or even other manual annotation tools. In some systems and methods, when dealing with labels/scores from imperfect sources (e.g., search engines), additional processes can be implemented to filter the initial labels and assess their reliability before using them as inputs for the propagation process.

Certain embodiments of the disclosed subject matter use the output of the brain signal analysis as an input to a label propagation system which propagates the initial brain-signal based labels to the remaining data and predict the most likely labels (or scores) for each data point on the graph or novel data not included in the graph.

The output of certain disclosed system embodiments can include refined or predicted labels (or scores indicating likelihood of positive detection) of some or all the images in the collection. These outputs can be used to identify additional positive samples matching targets of interest, which in turn can be used for a variety of functions, such as to train more robust classifiers, arrange the best presentation order for image browsing, or rearrange image presentations.

Certain embodiments of the disclosed subject matter use a refined or predicted label set to modify the initial set of data to be presented to a user in a brain signal based target detection system.

In a disclosed embodiment of a system and method in accordance with the disclosed subject matter, a partially labeled multimedia data set is received and an iterative graph-based optimization method is employed resulting in improved label propagation results and an updated data set with refined labels.

Embodiments of the disclosed systems and methods are able to handle label sets of unbalanced class size and weigh labeled samples based on their degrees of connectivity or other importance measures.

In certain embodiments of the disclosed methods and systems, after the propagation process is completed, the predicted labels of all the nodes of the graph can be used to determine the best order of presenting the results to the user. For example, the images can be ranked in the database in a descending order of likelihood so that user can quickly find additional relevant images. Alternatively, the most informative samples can be displayed to the user to obtain the user's feedback, so that the feedback and labels can be collected for those critical samples. These functions can be useful to maximize the utility of the user interaction so that the best prediction model and classification results can be obtained with the least amount of manual user input.

The graph propagation process can also be applied to predict labels for new data that is not yet included in the graph. Such processes can be based, for example, on nearest neighbor voting or some form of extrapolation from an existing graph to external nodes.

In some embodiments of the disclosed subject matter, to implement an interactive and real-time system and method, the graph based label propagation can use a graph superposition method to incrementally update the label propagation results, without needing to repeat computations associated with previously labeled samples.

BRIEF DESCRIPTION OF THE DRAWINGS

Further features and advantages of the presently disclosed subject matter will become apparent from the following detailed description taken in conjunction with the accompanying figures showing illustrative embodiments of the disclosed subject matter, in which:

FIG. 1 is a diagram illustrating exemplary aspects of computer vision system modes in accordance with the presently disclosed subject matter;

FIG. 2 is diagram illustrating an exemplary graphic user interface (GUI) portion of a computer vision module in accordance with the presently disclosed subject matter;

FIG. 3 is a flow chart illustrating an exemplary computer vision labeling propagation and refining method in accordance with the presently disclosed subject matter;

FIG. 4 is a diagram illustrating a fraction of a computer vision constructed graph and computation of a computer vision node regularizer method in accordance with the presently disclosed subject matter;

FIG. 5 is a flow chart illustrating an exemplary computer vision labeling diagnosis method in accordance with the presently disclosed subject matter.

FIG. 6 illustrates hardware and functional components of an exemplary system for brain signal acquisition and processing;

FIG. 7 is a diagram illustrating exemplary aspects of a combined brain-computer multimedia processing system in accordance with the presently disclosed subject matter;

FIG. 8 is a flow chart illustrating an exemplary method according to the presently disclosed subject matter.

FIGS. 9A-9G together form a diagram illustrating exemplary aspects of a virtual-world user-navigation system in accordance with the presently disclosed subject matter.

DETAILED DESCRIPTION

Systems and methods as disclosed herein can be used to overcome the labeling and classification deficiencies of prior systems and methods described above by coupling both computer vision and human vision components in various configurations. Computer vision components will first be described in accordance with the present disclosure.

FIG. 1 illustrates a system and various exemplary usage modes in accordance with the presently disclosed subject matter.

Given a collection of multimedia files, the exemplary computer vision components of FIG. 1 can be used to build an affinity graph to capture the relationship among individual images, video, or other multimedia data. One exemplary computer vision system can be a transductive annotation by graph (TAG) data processing system. The affinity between multimedia files can be represented graphically in various ways, for example: a continuous valued similarity measurement or logic associations (e.g., relevance or irrelevance) to a query target, or other constraints (e.g., images taken at the same location). The graph can also be used to propagate information from labeled data to unlabeled data in the same collection.

As illustrated in FIG. 1, each node in the graph 150 can represent a basic entity (data sample) for retrieval and annotation. In certain embodiments, nodes in the graph 150 can be associated with either a binary label (e.g., positive vs. negative) or a continuous-valued score approximating the likelihood of detecting a given target. The represented entity can be, for example, an image, a video clip, a multimedia document, or an object contained in an image or video. In an ingestion process, each data sample can first be pre-processed 120 (e.g., using operations such as scaling, partitioning, noise reduction, smoothing, quality enhancement, and other operations as are known in the art). Pre-filters can also be used to filter likely candidates of interest (e.g., images that are likely to contain targets of interest). After pre-processing and filtering, features can be extracted from each sample 130. TAG systems and methods in accordance with the disclosed subject matter do not necessarily require usage of any specific features. A variety of feature sets preferred by practical applications can be used. For example, feature sets can be global (e.g., color, texture, edge), local (e.g., local interest points), temporal (e.g. motion), and/or spatial (e.g., layout). Also, multiple types and modalities of features can be aggregated or combined. Given the extracted features, affinity (or similarity) between each pair of samples is computed 140. No specific metrics are required by TAG, though judicious choices of features and similarity metrics can significantly impact the quality of the final label prediction results. The pair-wise affinity values can then be assigned and used as weights of the corresponding edges in the graph 150. Weak edges with small weights can be pruned to reduce the complexity of the affinity graph 150. Alternatively, a fixed number of edges can be set for each node by finding a fixed number of nearest neighbors for each node.

Once the affinity graph 150 is created, a TAG system can be used for retrieval and annotation. A variety of modes and usages could be implemented in accordance with the teachings herein. Two possible modes include: interactive 160 and automatic 170 modes. In the Interactive Mode 160, users can browse, view, inspect, and label images or videos using a graphic user interface (GUI), an embodiment of which is described in more detail hereinafter in connection with FIG. 2.

Initially, before any label is assigned, a subset of default data can be displayed in the browsing window of the GUI based on, for example, certain metadata (e.g., time, ID, etc.) or a random sampling of the data collection. Using the GUI, a user can view an image of interest and then provide feedback about relevance of the result (e.g., marking the image as “relevant” or “irrelevant” or with multi-grade relevance labels). Such feedback can then be used to encode labels which are assigned to the corresponding nodes in the graph.

In Automatic Mode 170, the initial labels of a subset of nodes in the graph can be provided by external filters, classifiers, or ranking systems. For example, for a given target, an external classifier using image features and computer vision classification models can be used to predict whether the target is present in an image and assign the image to the most likely class (positive vs. negative or one of multiple classes). As another example, if the target of interest is a product image search for web-based images, external web image search engines can be used to retrieve most likely image results using a keyword search. The rank information of each returned image can then be used to estimate the likelihood of detecting the target in the image and approximate the class scores which can be assigned to the corresponding node in the graph. An initial label set can also be generated based on the initial output of the human vision components of FIG. 1.

In this particular embodiment, the TAG system hardware configuration can include an audio-visual (AV) terminal, which can be used to form, present or display audio-visual content. Such terminals can include (but are not limited to) end-user terminals equipped with a monitor screen and speakers, as well as server and mainframe computer facilities in which audio-visual information is processed. In such an AV terminal, desired functionality can be achieved using any combination of hardware, firmware or software, as would be understood by one of ordinary skill in the art. The system can also include input circuitry for receiving information to be processed. Information to be processed can be furnished to the terminal from a remote information source via a telecommunications channel, or it can be retrieved from a local archive, for example. The system further can include processor circuitry capable of processing the multimedia and related data and performing computational algorithms. The processor circuitry may be a microprocessor, such as those manufactured by Intel, or any other processing unit capable of performing the processing described herein. Additionally, the disclosed system can include computer memory comprising RAM, ROM, hard disk, cache memory, buffer memory, tape drive, or any other computer memory media capable of storing electronic data. Notably, the memory chosen in connection with an implementation of the claimed subject matter can be a single memory or multiple memories, and can be comprised of a single computer-readable medium or multiple different computer-readable media, as would be understood by one of ordinary skill in the art.

FIG. 2 shows an exemplary system GUI that can optionally be implemented in accordance with the presently disclosed subject matter. The disclosed GUI can include a variety of components. For example, image browsing area 210, as shown in the upper left corner of the GUI, can be provided to allow users to browse and label images and provide feedback about displayed images. During the incremental annotation procedure, the image browsing area can present the top ranked images from left to right and from top to bottom, or in any other fashion as would be advantageous depending on the particulars of the application. System status bar 220 can be used to display information about the prediction model used, the status of current propagation process and other helpful information. The system processing status as illustrated in FIG. 2 can provide system status descriptions such as, for example, ‘Ready’, ‘Updating’ or ‘Re-ranking.’ The top right area 230 of the GUI can be implemented to indicate the name of current target class, e.g., “statue of liberty” as shown in FIG. 3. For semantic targets that do not have prior definition, this field can be left blank or can be populated with general default text such as “target of interest.” Annotation function area 240 can be provided below the target name area 230. In this embodiment, a user can choose from labels such as ‘Positive’, ‘Negative’, and ‘Unlabeled.’ Also, statistical information, such as the number of positive, negative and unlabeled samples can be shown. The function button in this embodiment includes labels ‘Next Page’, ‘Previous Page’, ‘Model Update’, ‘Clear Annotation’, and ‘System Info.’

Various additional components and functions can be implemented. For example, image browsing functions can be implemented in connection with such a system and method. After reviewing the current ranking results or the initial ranking, in this embodiment, such functionality can be implemented to allow a user to browse additional images by clicking the buttons ‘Next Page’ and ‘Previous Page.’ Additionally, a user can also use the sliding bar to move through more pages at once.

Manual annotation functions can also optionally be implemented. In certain embodiments, after an annotation target is chosen, the user can annotate specific images by clicking on them. For example, in such a system, positive images can be marked with a check mark, negative images can be marked with a cross mark ‘x’, and unlabeled images can be marked with a circle ‘∘’.

Automatic propagation functions can also be implemented in connection with certain embodiments. After a user inputs some labels, clicking the button ‘Model Update’ can trigger the label propagation process and the system will thereafter automatically infer the labels and generate a refined ranking score for each image. A user can reset the system to its initial status by clicking the button labeled ‘Clear Annotation.’A user can also click the button labeled ‘System Info’ to generate system information, and output the ranking results in various formats that would be useful to one of ordinary skill in the art, such as, for example, a MATLAB-compatible format.

In the GUI embodiment shown in FIG. 2, two auxiliary functions are provided which are controlled by checking boxes ‘Instant Update’ and ‘Hide Labels.’ When a user selects ‘Instant Update,’ the shown system will respond to each individual labeling operation and instantly update the ranking list. The user can also hide the labeled images and only show the ranking results of unlabeled images by checking ‘Hide Labels.’

Given assigned labels or scores for some subset of the nodes in the graph (the subset is usually but not necessarily a small portion of the entire graph), embodiments of the disclosed systems can propagate the labels to other nodes in the graph accurately and efficiently.

FIG. 3 is a chart illustrating a TAG labeling propagation method in accordance with an exemplary implementation of the presently disclosed subject matter. At 310, the similarity or association relations between data samples are computed or acquired to construct an affinity graph. In 320, some graph quantities, including a propagation matrix and gradient coefficient matrix are computed based on the affinity graph. At 330, an initial label or score set over a subset of graph data is acquired. In various embodiments, this can be done via either interactive or automatic mode, or by some other mode implemented in connection with the disclosed subject matter. At 340, one or more new labels are selected and added to the label set. Procedure 350 is optional, where one or more unreliable labels are selected and removed from the existing label set. In 360, cleaned label set are obtained and a node regularization matrix is updated to handle the unbalanced class size problem of label data set. Procedures 340, 350, and 360 can be repeated until a certain number of iterations or some stop criteria are met. In step 370, the final classification function and prediction scores over the data samples are computed.

Additional description of computer vision algorithms and graph data generally described above is now provided. In an embodiment in accordance with the disclosed subject matter, an image set X=(X_L,X_U) can include labeled samples X_L={x₁, . . . , x_l} and unlabeled samples X_U={x_l+1, . . . , x_n }, where l is the number of labels. The corresponding labels for the labeled data set can be denoted as {y₁, . . . , y_l}, where yε{l, . . . , c} and c is the number of classes. For transductive learning, an objective is to infer the labels {y₁₊₁, . . . , y_n} of the unlabeled data X_U={x_l+1, . . . , x_n }, where typically l<<n, namely only a very small portion of data are labeled. Embodiments can define an undirected graph represented by G={X,E}, where the set of node or vertices is X={x_i} and the set of edges is E={e_ij}. Each sample x_i can be treated as the node on the graph and the weight of edge e_ij can be represented as w_ij. Typically, one uses a kernel function k(•) over pairs of points to calculate weights, in other words w_ij=k(x_i,x_j) with the RBF kernel being a popular choice. The weights for edges can be used to build a weight matrix which can be denoted by W={w_ij}. Similarly, the node degree matrix D=diag(d₁, . . . , d_n) can be defined as

$d_l=\sum _{j=l}^nW_{\mathrm{ij}}.$

A graph related quantity Δ=D−W is called graph Laplacian and its normalized version is

$L=D^{-\frac{1}{2}}\Delta D^{-\frac{1}{2}}=I-D^{-\frac{1}{2}}{\mathrm{WD}}^{-\frac{1}{2}}=I-S$

where

$S=D^{-\frac{1}{2}}{\mathrm{WD}}^{-\frac{1}{2}}.$

The binary label matrix Y can be described as YεB^n×c with Y_ij=1 if x_i has label y_i=j (means data x_i belongs to class j) and Y_ij=0 otherwise (means data x_i is unlabeled). A data sample can belong to multiple classes simultaneously and thus multiple elements in the same row of Y can be equal to 1. FIG. 4 shows a fraction of a representative constructed graph with weight matrix W, node degree matrix D, and label matrix Y. A classification function F, can then be estimated on the graph to minimize a cost function. The cost function typically enforces a tradeoff between the smoothness of the function over the graph and the accuracy of the function at fitting the label information for the labeled nodes.

Embodiments of the disclosed TAG systems and methods can provide improved quality of label propagation results. For example, disclosed embodiments can include: superposition law based incremental label propagation; a node regularizer for balancing label imbalance and weighting label importance; alternating minimization based label propagation; and label diagnosis through self tuning.

Embodiments of the disclosed TAG systems and methods can also include an incremental learning method that allows for efficient addition of newly labeled samples. Results can be quickly updated using a superposition process without repeating the computation associated with the labeled samples already used in the previous iterations of propagation. Contributions from the new labels can be easily added to update the final prediction results. Such incremental learning capabilities can be useful for achieving real-time responses to a user's interaction. Since the optimal prediction can be decomposed into a series of parallel problems, and the prediction score for individual class can be formulated as component terms that only depend on individual columns of a classification matrix F:

$F={\left(I-\alpha S\right)}^{-1}\sum _{i=1}^l{\hat{Y}}_i=\sum _{i=1}^l{\left(I-\alpha \hat{S}\right)}^{-1}Y_i=\sum _{i=1}^l{\hat{F}}_i$

where αε(0,1) is a constant parameter. Because each column of F encodes the label information of each individual class, such decomposition reveals that biases can arise if the input labels are disproportionately imbalanced. Prior propagation algorithms often failed in this unbalanced case, as the results tended to be biased towards the dominant class. To overcome this problem, embodiments disclosed herein can apply a graph regularization method to effectively address the class imbalance issue. Specifically, each class can be assigned an equal amount of weight and each member of a class can be assigned a weight (termed as node regularizer) proportional to its connection density and inversely proportional to the number of samples sharing the same class.

$F=\sum _{i=1}^l\widehat{v_{\mathrm{ii}}F_i}=\sum _{i=1}^l\widehat{{\left(I-\alpha S\right)}^{-1}v_{\mathrm{ii}}Y_i}={\left(I-\alpha S\right)}^{-1}\mathrm{VY}$

where the diagonal matrix V={v_ii} is introduced as a node regularizer to balance the influence of labels from different classes. Assume sample x_i is associated with label j, the value of v_ii is computed as:

$v_{\mathrm{ii}}=d_i/\sum _{k=1}^ld_ky_{\mathrm{kj}}$

where d_i is the node degree of labeled sample x_i and

$\sum _{k=1}^ld_kY_{\mathrm{kj}}$

is the sum of node degree of the labeled nodes in class j. FIG. 5 illustrates the calculation of node regularizer on a fraction of an exemplary constructed graph. The node weighting mechanism described above allows labeled nodes with a high degree to contribute more during the graph diffusion and label propagation process. However, the total diffusion of each class can be kept equal and normalized to be one. Therefore the influence of different classes can be balanced even if the given class labels are unbalanced. If class proportion information is known beforehand, it can be integrated into particular systems and methods by scaling the diffusion with the prior class proportion. Because of the nature of graph transduction and unknown class prior knowledge, however, equal class balancing leads to generally more reliable solutions than label proportional weighting.

Along with the node regularizer, incremental learning by superposition law is described here as another embodiment of the disclosed systems and methods. Let

$D_j=\sum _{k=1}^ld_kY_{\mathrm{kj}}$

denotes the total degree of the current labels in class j. Adding a new labeled sample x_s (the corresponding degree is d_ss) to class j, two coefficients λ, γ can be calculated as:

$\lambda =\frac{D_j}{D_j+d_{\mathrm{ss}}}$ $\gamma =\frac{d_{\mathrm{ss}}}{D_j+d_{\mathrm{ss}}}$

Then the new prediction score for class j can be rapidly computed as:

F _.j ^new =λF _.j +γP _.s

where F_.J is the j th column of the classification matrix F and P_.S is the j th column of the propagation matrix P (The propagation matrix will be defined later). This is in contrast to a brute force approach that uses the whole set of labeled samples, including the new labeled sample and the existing labeled samples, to calculate the classification function from scratch again. The disclosed systems and methods result in a much more efficient implementation of the label propagation process.

Certain embodiments of the disclosed systems and methods make modifications to the cost function used in previously used systems and methods. For example, in certain systems and methods, the optimization is explicitly shown over both the classification function F and the binary label matrix Y:

(F*,Y*)=arg min_{FεRn×c,YεBn×c}Q(F,Y)

where B is the set of all binary matrices Y of size n×c that satisfy Σ_jY_ij=1 for a single labeling problem, and for the labeled data x_iεX_l, Y_ij=1 if y_i=j. However, embodiments of the disclosed systems and methods naturally adapt to a multiple-label problem, where single multimedia file can be associated with multiple semantic tags. More specifically, the loss function is:

$Q\left(F,Y\right)=\frac{1}{2}\mathrm{tr}\left\{F^T\mathrm{LF}+{\mu \left(F-\mathrm{VY}\right)}^T\left(F-\mathrm{VY}\right)\right\}$

where the parameter μ balances two parts of the cost function. The node regularizer V permits the use of a normalized version of the label matrix Z defined as: Z=VY. By definition, in certain embodiments, the normalized label matrix satisfies Σ_izij=1.

An alternating minimization procedure to solve the above optimization problem can also contribute to improvements over prior methods and systems, as disclosed herein. Specifically, the cost function discussed above includes two variables that can be optimized. While simultaneously recovering both solutions can be difficult due to the mixed integer programming problem over binary Y and continuous F, a greedy alternating minimization approach can be used instead. The first update of the continuous classification function F is straightforward since the resulting cost function is convex and unconstrained, which allows the optimal F to be recovered by setting the partial derivative

$\frac{\partial Q}{\partial F}$

equal to zero. However, since YεB^n×cis a binary matrix and subject to certain linear constraints, another part of another embodiment of the disclosed alternating minimization requires solving a linearly constrained max cut problem which is NP. Due to the alternating minimization outer loop, investigating guaranteed approximation schemes to solve a constrained max cut problem for Y can be unjustified due to the solution's dependence on the dynamically varying classification function F during an alternating minimization procedure. Instead, embodiments of the currently disclosed methods and systems can use a greedy gradient-based approach to incrementally update Y while keeping the classification function F at the corresponding optimal setting. Moreover, because the node regularizer term V normalizes the labeled data, updates of V can be interleaved based on the revised Y.

The classification function, FεR^n×c, as used in certain embodiments of the disclosed subject matter, is continuous and its loss terms are convex, which allows its minimum to be recovered by zeroing the partial derivative:

$\frac{\partial Q}{\partial F}=0\Rightarrow \mathrm{LF}+\mu \left(F^{*}-\mathrm{VY}\right)=0\Rightarrow F^{*}={\left(L/\mu +I\right)}^{-1}\mathrm{VY}=\mathrm{PVY}$

where P=(L/μ+I)⁻¹/is denoted as the propagation matrix and can assume the graph is symmetrically built. To update Y, first Y can be replaced by its optimal value F* as shown in the equation above. Accordingly:

$\begin{array}{c}Q\left(Y\right)=\frac{1}{2}\mathrm{tr}\left(Y^TV^TP^T\mathrm{LPVY}+\mu {\left(\mathrm{PVY}-\mathrm{VY}\right)}^t\left(\mathrm{PVY}-\mathrm{VY}\right)\right)\\ =\frac{1}{2}\mathrm{tr}\left(Y^TV^T\left[P^T\mathrm{LP}+\mu \left(P^t-I\right)\left(P-I\right)\right]\mathrm{VY}\right)\end{array}$

This optimization still involves the node regularizer V, which depends on Y and normalizes the label matrix over columns. Due to the dependence on the current estimate of F and V, only an incremental step will be taken greedily in certain disclosed embodiments to reduce Q(Y). In each iteration, position (i*, j*) in the matrix Y can be found and the binary value Y_i*j* of can be changed from 0 to 1. The direction with the largest negative gradient can guide the choice of binary step on Y. Therefore,

$\frac{\partial Q}{\partial Y}$

can be evaluated and the associated largest negative value can be found to determine (i*, j*).

Note that setting Y_i*j*=1 is equivalent to a similar operation on the normalized label matrix Z by setting Z_i*j*=ε0<ε<1, and Y, Z to have one-to-one correspondence. Thus, the greedy minimization of Q with respect to Y in this disclosed embodiment is equivalent to the greedy minimization of Q with respect to Z:

$\left(i^{*},j^{*}\right)=\arg \underset{i,j}{\min }\frac{\partial Q}{\partial Z}$

The loss function can be rewritten using the variable Z as:

$Q\left(Z\right)=\frac{1}{2}\mathrm{tr}\left(Z^T\left[P^T\mathrm{LP}+\mu \left(P^T-I\right)\left(P-I\right)\right]Z\right)=\frac{1}{2}\mathrm{tr}\left(Z^T\mathrm{AZ}\right)$

where A represents A=P^T LP+μ(P^T−I)(P−I). Note that A is symmetric if the graph is symmetrically built. The gradient of the above loss function can be derived and recovered with respect to Z as:

$\frac{\partial Q}{\partial Z}=\mathrm{AZ}=\mathrm{AVY}.$

As described earlier, the gradient matrix can be searched to find the minimal element for updating the following equation:

(i*,j*)=arg min_{xεXu,1≦j≦c}∇Z_ijQ

The label matrix can be updated by setting Y_i*j*=1. Because of the binary nature of Y, Y_i*j* can be set to equal 1 instead of using a continuous gradient approach. Accordingly, after each iteration, the node regularizer can be recalculated using the updated label matrix.

The updated Y in accordance with certain disclosed embodiments is greedy and could therefore oscillate and backtrack from predicted labeling in previous iterations without convergence guarantees. To guarantee convergence and avoid backtracking, inconsistency or unstable oscillation in the greedy propagation of labels, in preferred embodiments, once an unlabeled point has been labeled, its labeling can no longer be changed. In other words, the most recently labeled point (i*, j*) is removed from future consideration and the algorithm only searches for the minimal gradient entries corresponding to the remaining unlabeled samples. Thus, to avoid changing the labeling of previous predictions, the new labeled node x_i can be removed from X_u and added to X_l.

The following equations summarize the updating rules from step t to t+1 in certain embodiments of the scheme of graph transduction via alternative minimization (GTAM). Although the optimal F* can be computed in each iteration, it does not need to explicitly be updated. Instead, it can be implicitly used to directly updated Y:

${\nabla }_ZQ^t=A·V^tY^t\left(i^{*},j^{*}\right)=\underset{x_i\in X_u,1\le j\le c}{\mathrm{argmin}}{\nabla }_{Z_{\mathrm{ij}}}Q^t$ $Y_{i^{*}j^{*}}^{t+1}=1$ $v_{\mathrm{ii}}^{t+1}=d_i/\sum _{k=1}^ld_kY_{\mathrm{kj}}^{t+1}$ $X_U^{t+1}\leftarrow X_L^t+x_{i^{*}};X_U^{t+1}\leftarrow X_U^t-x_{i^{*}}$

The procedure above can repeat until all points have been labeled in connection with the label propagation of the disclosed subject matter.

To handle errors in a label set, embodiments of the disclosed methods and systems can be extended to formulate a graph transduction procedure with the ability to handle mislabeled instances. A bidirectional greedy search approach can be used to simultaneously drive wrong label correction and new label inferencing. This mechanism can allow for automatic pruning of incorrect labels and maintain a set of consistent and informative labels. Modified embodiments of the systems and methods disclosed earlier can be equipped to more effectively deal with mislabeled samples and develop new “Label Diagnosis through Self Tuning” (LDST) systems and methods. In an exemplary embodiment of these systems and methods, a set of initial labels is acquired. They can be acquired, for example, either by user annotation or from another resource, such as text-based multimedia search results. The gradient of the cost function with respect to label variable is computed based on the current label set is computed, and a label is added from said unlabeled data set based on the greedy search, i.e., finding the unlabeled sample with minimum gradient value. A label is then removed from said label set based on the greedy search, i.e., finding the labeled sample with maximum gradient value. The last two can be performed in reverse order without losing generalization, and can be executed a variable number of times (e.g., several new labels can be added after removing an existing label). Certain embodiments of the disclosed systems and methods update the computed gradients based on the new label set and repeat the last two parts of the procedure to retrieve a refined label set.

Embodiments of the disclosed LDST systems and methods can be used to improve the results of text based image search results. In one embodiment, top-ranked images can be truncated to create a set of pseudo-positive labels, while lower-ranked images can be treated as unlabeled samples. LDST systems and methods can then be applied to tune the imperfect labels and further refine the rank list. Additional embodiments can be used on a variety of data set types, including text classification on webpages and to correctly identify handwritten data samples.

The human vision components of the presently disclosed subject matter (including the brain-computer interface) are now described.

As previously mentioned, in recent years, there has been substantial interest in decoding the human brain state. There have been a variety of neural signals which have been targeted for decoding, ranging from spike trains collected via invasive recordings to hemodynamic changes measured via non-invasive fMRI. Systems and methods in accordance with the disclosed subject matter use EEG as a non-invasive measure to relate brain state to events correlated with the detection of “interesting” visual objects and images. One example of a robust signal using EEG is the P300. The P300 reflects a perceptual orienting response or shift of attention which can be driven by the content of the sensory input stream. Oscillatory brain activity often found during resting state (10 Hz oscillations known as “alpha” activity) as well as transient oscillations sometimes associated with perceptual processing (30 Hz and higher known as “gamma” activity) can also be indicative of a subject's attention state.

Certain systems and methods in accordance with the disclosed subject matter distinguish between two distinct brain states: (+) positive states in which the subject sees something of interest in an image, versus (−) negative states for which the image contains nothing of particular interest. This distinction is not to deduce from the brain signal what the exact content is or what the subject sees in the image, but instead, to utilize the high temporal resolution of EEG to detect individual recognition events from just a short segment of EEG data. For individual images that are presented to a subject for as little as 100 ms, exemplary embodiments of the disclosed systems and methods can detect the brain signals elicited by positive examples, and distinguish them from the brain activity generated by negative example images. A task for exemplary disclosed analysis systems and methods is to classify the signal between two possible alternatives.

In exemplary systems and methods in accordance with the disclosed subject matter, a fast image sequences is presented to a user via a process known as rapid serial visual presentation (RSVP). In exemplary embodiments of RSVP, images can be presented very rapidly, for example, at rates of 5 to 10 images per second. To classify brain activity elicited by these images, certain exemplary methods analyze 1 s of data, recorded with multiple surface electrodes, following the presentation of an image.

Systems and methods in accordance with the disclosed subject matter can be used to measure linear variations in EEG measurement signals. By averaging over EEG measurements with appropriate coefficients (positive or negative with magnitudes corresponding to how discriminant each electrode is) it is possible to obtain a weighted average of the electrical potentials that can be used to differentiate positive from negative examples as represented below:

$y_t=\sum _iw_ix_{\mathrm{it}}$

Here x_it represents the electrical potential measured at time t for electrode i on the scalp surface, while w_i represents the spatial weights which have to be chosen appropriately. A goal of this summation is to combine voltages linearly such that the sum y is maximally different between two conditions. This can be thought of as computing a neuronal current source y_t that differs most between times samples t+ following positive examples and the times t− following negative examples, y_t+>y_t−. Label ‘+’ indicates that the expression is evaluated with a signal x_it recorded following positive examples and label ‘−’ indicates the same for negative examples. There are a number of algorithms available to find the optimal coefficients w_i in such a binary linear classification problem, e.g., Fisher Linear Discriminants (FLD), Penalized Logistic Regression (PLR), and Support Vectors Machines (SVM).

In certain methods and systems in accordance with the disclosed subject matter, optimal weight vectors, w_ki are calculated for a time window following the presentation of the data (index k labels the time window):

$y_{\mathrm{kt}}=\sum _iw_{\mathrm{ki}}x_{\mathrm{it}},t=\left(k-1\right)T\dots \mathrm{kT}$

These different current sources y_kt can then be combined in an average over time to provide the optimal discriminant activity over the relevant time period:

$y=\sum _t\sum _kv_ky_{\mathrm{tk}};$

For an efficient on-line implementation of this method, FLD can be used to train coefficients w_ik within each window of time, i.e., w_ik is trained such that y_kt+>y_kt−. The coefficients v_k can be used learned using PLR after all exemplars have been observed such that y₊>y₋. Because of the two-part process of first combining activity in space, and then again in time, this algorithm can be referred to as a “Hierarchical Discriminant Component Analysis”

Note that the first part in the exemplary embodiments described above does not average over time samples within each window. Instead, each time sample provides a separate exemplar that is used when training the FLD. For instance, a system with 50 training exemplars and 10 samples per window results in 500 training samples for a classification algorithm that can need to find 64 spatial weighting coefficients w_ik for the kth window. These multiple samples within a time window will correspond to a single examplar image and are therefore not independent. They do, however, provide valuable information on the noise-statistic: variations in the signal within the time window are assumed to reflect non-discriminant “noise.” In other words, one can assume that spatial correlation in the high-frequency activity (f>1/T) is shared by the low-frequency discriminant activity. In addition, by training the spatial weight separately for each window one assume that the discriminant activity is not correlated in time beyond the time window time scale. Both these assumptions contribute to a system's ability to combine thousands of dimensions optimally despite the small number of known training images.

The method described above combines activity linearly. This is motivated by the notion that a linear combination of voltages corresponds to a current source, presumably of neuronal origin within the skull. Thus, this type of linear analysis is sometimes called source-space analysis (“beam-forming” is a common misnomer for the same). A general form of combining voltages linearly in space and time can be represented as:

$y=\sum _t\sum _iw_{\mathrm{it}}x_{\mathrm{it}}$

However, the number of free parameters w_it it in this general form is the full set of dimensions—6,400 for certain embodiments, for example—with only a handful of positive exemplars to choose their values. To limit the degrees of freedom one can restrict the matrix w_it it to be of lower rank, e.g., K. The linear summation can then be written as:

$y=\sum _t\sum _iw_{\mathrm{it}}x_{\mathrm{it}},$

where is a low-rank bilinear representation of the full parameter space.

This bilinear model assumes that discriminant current sources are static in space with their magnitude (and possibly polarity) changing in time. The model allows for K such components with their spatial distribution captured by u_ik and their temporal trajectory integrated with weights v_tk. Again, a goal is to find coefficient u_ik, v_tk such that the bilinear projection is larger for positive examples than for negative examples, i.e., y₊>y₋. Notably, the x values referenced above need not be a time-domain signal, but could also be in the frequency domain. The linear integration could be performed in either domain.

The algorithms presented so far capture a type of activity that is often referred to as event related potentials (ERP). This term, ERP, refers to activity that is evoked in a fixed temporal relationship to an external event, that is, positive and negative deflections occur at the same time relative to the event—for example, the time of image presentation. In addition to this type of evoked response activity the EEG often shows variations in the strength of oscillatory activity. Observable events can change the magnitude of ongoing oscillatory activity or can induce oscillations in the EEG. To capture the strength of an oscillation, irrespective of polarity, it is common to measure the “power,” or the square of the signal, typically after it has been filtered in a specific frequency band. Instead of a linear combination, to capture power, one has to allow for a quadratic combination of the electrical potentials.

Once an interest score, y, is calculated, it can be converted to an interest label for use in a multimedia analysis/computer vision system as described above. More specifically, a binarization function g(•) can be applied to convert interest scores to multimedia labels as y=g(e), where yiε{1,0} and yi=1 for ei>ε, otherwise yi=0. The value ε can be referred to as an interest level for discretizing the EEG scores

Using the principles of computer vision systems (such as TAG) and brain-computer interfaces (EEG-based) as set forth above, systems and methods according to the disclosed subject matter herein can be implemented.

In accordance with the disclosed subject matter, it is possible to implement computer vision followed by EEG-RSVP systems and methods described above. Given prior information of a target type which can be instantiated in a TAG-based model, including contextual cues, TAG-based processing can operate on a dataset so as to eliminate regions of very low target probability and also provide an initial ordering of regions having high target probability. In addition, TAG can center image chips of potential regions of interest (ROIs) in a large image or set of images, which improves detection since potential targets are foveated when images are presented to subject. The top M images of the reordered dataset, in which sensitivity is high but specificity can be low, can be sampled and presented to the subject for EEG-RSVP analysis. The brain-computer interface processing can be tuned to produce high sensitivity and low specificity, with the EEG-RSVP mode can be used to increase specificity while maintaining sensitivity.

FIG. 6 illustrates the hardware components of a particular embodiment of a subsystem for brain data acquisition in accordance with the disclosed subject matter. Such a subsystem can include EEG electrodes 610, which may be passive or active electrodes. The electrodes are connected to an EEG amplifier 620, which processes and amplifies the EEG signals for further analysis. An analog-to-digital converter 630 is then used to input the data received from the amplifier into a computer 650. Interface 640 between the A-D converter 630 and the computer 650 may be a wire interface connected via USB or other standard, or a wireless connection via Bluetooth or other standard, or any other known mechanism for data transfer. In certain embodiments, the system hardware implementation can use multiple computers 650, such as three personal computers (laptop, desktop, handheld, or any other personal computing device), two used for the RSVP and EEG recording and classification, and one for image processing, or the functionality of all modules could be performed from a single computer. One of ordinary skill in the art would understand a variety of different configurations of such a system, including a general purpose personal computer programmed with software sufficient to enable the methods of the disclosed subject matter described herein.

In one exemplary embodiment, the analysis system utilizes a 64 electrode EEG recording system in a standard montage. EEG can be recorded at, for example, a 1 kHz sampling rate. While the EEG is being recorded, the RSVP display module can uses a dedicated interface to display blocks of images at the specified frame rate. In certain embodiments, blocks are typically 100 images long. The frame rate can be set to 5 or 10 Hz depending on the target/imagery types, and the human observer's response to preliminary presentations. The interface draws from a pool of potential target chips and a pool of “distracters.” One role of the distracters is to achieve a desired prevalence of target chips, that will maintain the human observer engaged in the presentation: if the prevalence is too low or too high, the observer can not keep an adequate focus and can more easily miss detections. Given that the computer vision outputs include some false positives, the number of distracters used depends in fact on the expected number of true target chips from the computer vision module.

The exemplary EEG analysis module can receive a list of image chips and detection details from the computer vision module, that includes pixel locations and detection confidence scores, and uses this input to generate the RSVP image sequences that will be used for presentation to the subject and analysis. It then performs several tasks: it acquires and records the EEG signals, using for example the hardware identified in FIG. 6, orchestrates the RSVP, matches the EEG recordings with the presented images, trains an underlying classifier using training sequences, and uses the classifier with new generated image sequences.

In certain embodiments, a classification module relies on a hierarchical discriminant component analysis algorithm. At the first level, the classifier can use multiple temporal linear discriminators, each trained on a different time window relative to the image onset, to estimate EEG signatures of target detection. At a second level, the classifier can estimate a set of spatial coefficients that will optimally combine the outputs of the temporal discriminators to yield the final classification outcomes. The classification module is used in two different stages: training and actual usage with new imagery. The training can include a presentation of blocks with a set number of known targets in each block. The training sequences need not be related to the test sequences, in terms of their content, as the premise of the approach is that it detects objects of interest, but is not sensitive to the signatures of specific objects, and can therefore maintain its detection performance from one type of imagery to another.

Once an exemplary human imaging module has analyzed the relevant data, it can generate a list of images or image chips and their associated classification confidences, which can be used to prioritize the visualization of the corresponding images or image locations. The visualization interface permits the visualization of the prioritized locations in an adequate software environment for the task or user at hand. For example, for image analysts, certain embodiments use an interface to RemoteView, an imagery exploitation software application often used in the GeoIntelligence community. The interface provides a play control like toolbar that lets the analyst jump from one prioritized location to the next, while the analyst still retains access to all of RemoteView's functionality.

It is also possible and useful in connection with the disclosed subject matter to implement EEG-RSVP analysis systems and methods followed by computer vision (such as TAG) systems and methods as part of a combined system and method. In certain exemplary embodiments, in the absence of prior knowledge of the target type or a model of what an “interesting” image is, the EEG-RSVP is first run on samples of D_i which can result in an image reordering in which images are ranked based on how they attracted the human subject's attention. This reordering can be used to generate labels for a computer vision based learning system which, given a partial labeling of D_i propagates these labels and re-orders the database. In this embodiment, EEG scores are numbers with more positive scores indicating that the subject was interested in or strongly attending to the presented multimedia data. The scores are sorted and the multimedia data associated with the top N scores are considered positives and labeled as such (given label +1) and used as training data for the TAG system. N can be chosen to be fixed (e.g., top 20 scores) or can be selected based on the requirements of a certain precision (e.g., the N scores where at least X % are true positives and 100−X % are false positives). In another embodiment, the lowest M ranked EEG scores are considered negatives and labeled as such (label=−1) so that N positive examples and M negative examples are provided to the TAG system. In another embodiment, the real-number values of the EEG scores are used to weight the strength of the training examples. For instance a EEG score of 0.3 for a EEG labeled image would result in a training label that is three times stronger than an EEG labeled image with a score of 0.1. The EEG-RSVP is designed to identify a small number “interesting” images which are then used by a semi-supervised computer vision system to reorder the entire image database.

Exemplary systems in which computer vision analysis follows EEG analysis can be similar to the alternative computer vision followed by EEG systems, in that they can use the same type of components, such as a computer vision module, an EEG analysis module, and a visualization/review module. However, in these systems, the EEG analysis module precedes the computer vision TAG components. Additionally, the number of examples provided by the EEG analysis can be insufficient to train conventional supervised learning algorithms, and there can be inaccuracies in the EEG outputs due to typically lower sensitivity of the approach. Therefore, a computer vision TAG module underpinned by a semi-supervised learning algorithm can be used to improve the EEG output. In certain embodiments, the outputs of the EEG systems and methods are a set of positive and negative examples (as determined by a suitable EEG confidence threshold), that can serve as labeled inputs to a graph-based classifier to predict the labels of remaining unlabeled examples in a database. TAG can then incorporate its automated graph-based label propagation methods and in real or near-real time generate refined labels for all remaining unlabeled data in the collection.

Further, as previously mentioned, it is also possible to have tightly coupled EEG systems and computer vision systems and methods in accordance with the disclosed subject matter. In certain embodiments, both EEG analysis and computer visions are run in parallel and coupled either at the feature space level, or at the output (or confidence measure) level, leading to a single combined confidence measure that can serve as a priority indicator. This mode can require prior information on the target type. These modes can also potentially include feedback or multiple iterations within a closed-loop system.

Additionally, in certain embodiments of the disclosed subject matter, data is first analyzed using a EEG system and the results are used as the basis for input to a computer vision system such as TAG. The computer vision system can then be used to refine the data to be presented to a human once again as part of further EEG-based analysis. Among other things, a closed-loop implementation such as this can allow for more refined results and more efficient analysis of large sets of data.

FIG. 7 is a diagram illustrating exemplary aspects of a combined human-computer/multimedia-processing system in accordance with the presently disclosed subject matter. In this particular embodiment, EEG-based generic interest detector 710 includes an interest object detector 710 which performs calculations resulting in an initial label (annotation) set 730. This set can be used as an input to a computer vision system 740 including, among other things, a visual similarity graph 750 and a label refinement module 760 which can be used to generate a refined label/annotation set 770.

FIG. 8 is a flow chart illustrating a combined EEG-based interest detection method coupled to a TAG labeling propagation method in accordance with an exemplary implementation of the presently disclosed subject matter. At 810, a system presents multimedia data to a user. At 820, the system receives user response data based on the user's brain signal response to the presented data. At 830, the system determines user interest in the presented data based on the EEG response data. 820 and 830 involve using hardware such as the system of FIG. 6 to, more specifically, receive EEG signals and amplify, decode, and process them in real time as the human subject is viewing a rapid succession of images being presented on a display. At 840, the system can extract relevant features from image data and generates an initial label set. At 850, the similarity or association relations between data samples are computed or acquired to construct an affinity graph. At 860, some graph quantities, including a propagation matrix and gradient coefficient matrix, are computed based on the affinity graph. At 870, an initial label or score set over a subset of graph data is acquired from the system component which generated the initial label set based on user response data. 850 and 860 can be performed before or after 870. At 875, which is optional, one or more new labels are selected and added to the label set. 880 is optional, wherein one or more unreliable labels are selected and removed from the existing label set. At 890, cleaned label set are obtained and a node regularization matrix is updated to handle the unbalanced class size problem of label data set. Note that 875, 880 and 890 can be repeated, if necessary, until a certain number of iterations or some stop criteria are met. At 895, the final classification function and prediction scores over the data samples are computed. The output of 895 can be used to select and arrange certain presentations of images to users as input to 810 to complete a looped integration system.

The described methods and systems can be applied to recognize, propagate, and recommend other types of information relevant to a user's interest. This functionality may also be implemented in other applications besides multimedia data labeling and retrieval.

For instance, certain embodiments of the disclosed systems and methods can also be used for web search improvements. Images on web sharing sites often are already associated with textual tags, assigned by users who upload the images. However, it is well known to those skilled in the art that such manually assigned tags can often be erratic and inaccurate. Discrepancies can be due, for example, to the ambiguity and/or non-uniformity of labels or lack of control of the labeling process. Embodiments of the disclosed systems and methods can be used to efficiently refine the accuracy of the labels and improve the overall usefulness of search results from these types of internet websites, and more generally, to improve the usefulness and accuracy of internet multimedia searches overall.

In embodiments involving user's interaction with online media objects, such as video, music, and social network website postings, analysis of a user's neural signatures can be used to produce the initial measures of user interest with respect to the various online media objects. The interest measures can then be used as initial labels which are then propagated over a graph to other objects. In a manner similar to that described above, the nodes in the graph can represent the media objects and the edges in the graph capture relationships among media objects, such as the similarity of videos or music, similarity of authors and genres, and similarity of the comments made by online users on a social network or other similar website. After the label propagation progress, predicted labels of other nodes can be used to estimate the potential interest of the user with respect to other media objects, which are then recommended to the user based on the sorted rankings of the predicted interest scores.

Certain exemplary embodiments of the presently disclosed subject matter can utilize these neural signatures to tag information in the virtual world so as to optimize navigation through the virtual world. Such navigation functions include, for example and without limitation, implementations found in electronic games, virtual reality, and augmented reality systems. The virtual world can be represented via a graph with nodes being locations, objects, or events in the world and edges being relationships between those locations, objects, or events. Each node has features that represent attributes of the objects or events, and context at that specific location. Node connectedness and features can be based on coordinates in the virtual world or similarity of the attributes (e.g., similar visual appearances or similar event categories). Neural signatures, measured for example via EEG, can be used as labels that reflect the subjective perceptual/cognitive/emotional state of the user given their location or viewing of the objects or events. Such neural signatures could reflect subjective states such has “how interesting is this location/object”, “how excited am I to be at this location or see this object/event”, as well as brain states that can be decoded via EEG or other neural measures. The resulting labels (or neural scores) can then be propagated through the graph-based representation of the virtual world to identify other locations, objects, or events in the world that would likely yield a similar perceptual/cognitive/emotional state for the user. This information can then be used in a variety of ways, for example, to make recommendations or plan a path for the user to visit specific locations, objects or events in the virtual world. The virtual world can be one in which the subject is an avatar in an otherwise rich multisensory world, a mobile user exploring a physical environment, or a participant in a more abstract world such as a social network.

FIG. 9 is a diagram illustrating exemplary aspects of a system for user navigation in an example virtual world in accordance with the presently disclosed subject matter. FIG. 9A shows an exemplary system GUI that can optionally be implemented in accordance with the presently disclosed subject matter. The GUI includes a situation awareness (SA) map 910 indicating the importance of locations in the virtual world based on the prediction of the user interest and possibly other prior knowledge; and the interface 920 for the user (as an avatar) to visualize the locations, objects, and events in the world. FIG. 9B illustrates the user interface 920 and the user 930 wearing a neural signal sensing device 940. FIG. 9C shows a higher-resolution version of the SA map 910, which indicates the importance of locations included in the world. FIG. 9D shows the graph representation of the virtual world. FIG. 9E shows the neural signals recorded by the device during the navigation process. FIG. 9F plots the initial measured interest scores on the graph based on the neural signal analysis. FIG. 9G shows the predicted labels after label propagation indicating the interest scores (importance) of all nodes.

Certain embodiments in connection with the disclosed subject matter can similarly be applied to use neural signatures for tagging information in the real world so as to optimize navigation and interaction with real world places and objects. Such navigation functions may include, without limitation, implementations found in mobile navigation systems. The real world can be represented via a graph with nodes being locations, objects, or events in the world and edges being relationships between those locations, objects, or events. Each node has features that represent attributes of the objects or events, and can include context at the specific location. Node connectedness and features can be based on geospatial position system (GPS) coordinates or similarity of the attributes (e.g., similar visual appearances or similar event categories). Neural signatures, measured for example via EEG, can be used as labels that reflect the subjective perceptual/cognitive/emotional state of the user given their location or viewing of the objects or events. As in the virtual world context above, such neural signatures could reflect subjective states such has “how interesting is this location/object”, “how excited am I to be at this location or see this object/event”, as well as brain states that can be decoded via EEG or other neural measures. The resulting labels (aka, neural scores) can then be propagated through the graph-based representation of the real world to identify other locations, objects, or events in the world that would likely yield a similar perceptual/cognitive/emotional state for the user. This information can then used in a variety of ways, for example, to make recommendations or plan a path for the user to visit specific locations, objects or events in the real world.

In accordance with the presently disclosed subject matter, the term “multimedia” may broadly encompass all manner of sensory data processed by human physical sensory systems, including, without limitation, traditional multimedia objects viewed on a display, such as video and images, audio, text; virtual-space objects such as those encountered in electronic games, virtual reality, and augmented reality; and even physical-space real-world objects such as locations, objects, events, sights, sounds, tastes, smells, and tactile sensations. This listing of examples is not intended to be exhaustive. In accordance with the disclosed subject matter, a user's brain states can be decoded via EEG to ascertain reactions to all manner of stimulus—from perception of on-screen displayed information to perception of experiences of places or objects in the real world—to achieve benefits in connection with the labeling, categorizing, searching and recommending applications disclosed in connection with the exemplary embodiments.

Because the disclosed systems and methods are scalable in terms of feature representation, other application specified features can also be utilized to improve the graph propagation.

In another embodiment, a system and method for collaborative search can be employed. In such an embodiment, multiple EEG scores can be received from multiple human observers simultaneously, with the human observers each being presented the same multimedia data. The multiple scores can be processed and used to construct labels for the displayed multimedia data.

Further, computer vision systems and methods for use with brain-computer interfaces and methods as described herein are not limited to TAG or LDST systems and methods. Other back-end systems that may be used to process the brain-computer interface data may include any type of statistical model, including clustering, support vector machines, belief networks, and kernel-based systems and methods. Any computer vision label propagation component may be utilized. Ultimately, the described brain-computer interface can be used in conjunction with a number of different computer analysis systems to achieve the principles of the disclosed subject matter.

The foregoing merely illustrates the principles of the disclosed subject matter. Various modifications and alterations to the described embodiments will be apparent to those skilled in the art in view of the teachings herein. Further, it should be noted that the language used in the specification has been principally selected for readability and instructional purposes, and can not have been selected to delineate or circumscribe the inventive subject matter. Accordingly, the disclosure herein is intended to be illustrative, but not limiting, of the scope of the disclosed subject matter, which is set forth in the following claims.

Claims

1. A computer-based method for labeling one or more media objects related to one or more products, comprising: presenting to a user a first of said one or more media objects corresponding to a first of said one or more products;receiving a user response from said user corresponding to said first media object;generating, using a processor, an object label corresponding to said first media object based at least in part on said user response, said object label comprising an interest score for the first media object; andassociating said object label for said first media object with said user.

2. The method of claim 1, further comprising: performing at least one of refining the object label associated with said first media object or predicting a new object label pertaining to said stored media objects by calculating a classification function; andpresenting to said user a second of said one or more media objects corresponding to a second of said one or more products, said second media object selected based on said interest score of said first product.

3. The method of claim 2, further comprising: receiving a second user response from said user corresponding to said second media object;generating, using a processor, a second object label corresponding to said second media object based at least in part on said second user response, said second object label comprising a second interest score for said second media object; andassociating said second object label for said second media object with said user.

4. The method of claim 1, further comprising: storing a media affinity graph in one or more memories, wherein said affinity graph represents media object samples as nodes and comprises edges measuring relatedness among said object samples; andcalculating a classification function based on at least said selected object label using a processor associated with said one or more memories, by iteratively performing at least updating said selected object label relating to said selected media object or predicting a new object label for said stored media objects.

5. The method of claim 1, wherein said user response comprises brain signal response data of said user.

6. The method of claim 5, wherein said generating comprises decoding said brain signal response data of said user.

7. The method of claim. 1, wherein said user response comprises an interaction with a graphical user interface.

8. The method of claim 1, wherein said first media object comprises image data.

9. The method of claim 1, wherein said first media object comprises video data.

10. The method of claim 1, wherein said first media object comprises audio data.

11. The method of claim 2, wherein predicting said new object label comprises automatically selecting a most informative media object, predicting its corresponding class and labeling the corresponding media object.

12. The method of claim 2, wherein said refining comprises performing a greedy search among the gradient direction of the classification function.

13. The method of claim 1, further comprising presenting said selected media object to said user one or more additional times and receiving a further user response to said selected media object one or more additional times to further refine said object label.

14. The method of claim 1, further comprising presenting navigation data to said user corresponding to said first product if said interest score exceeds a threshold.

15. A computer-based system for labeling one or more media objects related to one or more products, comprising: one or more memories;one or more processors coupled to said one or more memories, wherein said one or more processors are configured to:present to a user a first of said one or more media objects corresponding to a first of said one or more products;receive a user response from said user corresponding to said first media object;generate an object label corresponding to said first media object based at least in part on said user response, said object label comprising an interest score for the first media object; andassociate said object label for said first media object with said user.

16. The system of claim 15, wherein said one or more processors are further configured to: perform at least one of refining the object label associated with said first media object or predicting a new object label pertaining to said stored media objects by calculating a classification function; andpresent to said user a second of said one or more media objects corresponding to a second of said one or more products, said second media object selected based on said interest score of said first product.

17. The system of claim 16, wherein said one or more processors are further configured to: receive a second user response from said user corresponding to said second media object;generate a second object label corresponding to said second media object based at least in part on said second user response, said second object label comprising a second interest score for said second media object; andassociate said second object label for said second media object with said user.

18. The system of claim 15, wherein said one or more processors are further configured to: store a media affinity graph in one or more memories, wherein said affinity graph represents media object samples as nodes and comprises edges measuring relatedness among said object samples; andcalculate a classification function based on at least said selected object label using a processor associated with said one or more memories, by iteratively performing at least updating said selected object label relating to said selected media object or predicting a new object label for said stored media objects.

19. The system of claim 15, wherein said response comprises brain signal response data of said user.

20. The system of claim 19, wherein said generating comprises decoding said brain signal response data of said user.

21. The system of claim 15, wherein said user response comprises an interaction with a graphical user interface.

22. The system of claim 15, wherein said first media object comprises image data.

23. The system of claim 15, wherein said first media object comprises video data.

24. The system of claim 15, wherein said first media object comprises audio data.

25. The system of claim 16, wherein said one or more processors are further configured to predict said new object label at least in part by automatically selecting a most informative media object, predicting its corresponding class and labeling the corresponding media object.

26. The system of claim 16, wherein said one or more processors are further configured to refine at least in part by performing a greedy search among the gradient direction of the classification function.

27. The system of claim 15, wherein said one or more processors are further configured to present said selected media object to said user one or more additional times and receive a further user response to said selected media object one or more additional times to further refine said object label.

28. The system of claim 15, wherein the one or more processors are further configured to present navigation data to said user corresponding to said first product if said interest score exceeds a threshold.