Research Interests

Scalable Learning

Since different humans can perform the same actions differently hence, the activity recognition systems generally need a substantial amount of training data. Although TWSVM based classifiers are faster than conventional SVMs, these are not able to scale up to handle very large number of data samples during training as solving the corresponding quadratic optimization problem become infeasible. In contemplation of this drawback, working on the lines of Stochastic Gradient Support Vector Machine (SG-SVM) and Stochastic Gradient Twin Support Vector Machine (SG-TSVM) [3] which deals with training samples individually, we propose an efficient stochastic Quasi-Newton method based Twin Parametric Support Vector Machine termed as SQN-PTWSVM. As many studies reveal that Hinge loss function does not only make the problem non-smooth leading to instability near the minima but also propagate noise-sensitivity to the problem which makes the stochastic optimization process deviate near the problem which ultimately speeds up the training process as well point of non-differentiability, we propose to use Pinball loss function to minimize the training loss, which is also efficient and more robust to the presence of noise when compared to conventional hinge loss SVM for large-scale data scenarios. The resulting SQN-PTWSVM can effectively handle large training data in HAR framework. Unlike SG-TSVM, which has the limited competence to handle feature noise, SQN-PTWSVM is insensitive to noise, robust to re-sampling and leads to rapid convergence of the corresponding convex optimization problem which ultimately speeds up the training process as well.

Noise insensitive Learning

As mentioned above due to the excellent capability of Pinball error loss in handling feature noise present in data which eventually leads to robust classification models, it has attracted wide research attention recently. However, the main focus to ensure robustness to noise and outliers has led to limited sparsity in the classifier(s) which result in the increased testing time. In HAR framework, large testing time hinders the applicability of these systems in real-world scenarios. Thus, in our subsequent work, we develop two robust Pin-TWSVM based models termed as ep−TWSVM and Flex−Pin −TWSVM that not only handles the noise present in data but also control model sparsity using user regulated and self-optimized insensitive zone respectively. The insensitive zone obtained in the model corresponds to the zero value of the dual variable and thus lead to simpler and sparser model. Extensive experiments performed on standard machine learning datasets confirm the efficacy of the proposed noise-insensitive models which also presents better sparseness than related traditional classifiers while giving comparable prediction performance.

Semi-supervised Learning

Human activity recognition systems, like many other machine learning problems such as Content-Based Image Retrieval Systems, usually poses yet another interesting challenge. Although the acquisition of hundreds of hours of training videos is indeed an easy task but obtaining corresponding training label is difficult, tedious or sometimes just boring. In such cases, usually semi-supervised learning approaches such as Laplacian SVM, requiring limited label information, are employed. However, conventional semi-supervised learning approaches often have no explicit control over the choice or usefulness of labeled data available for training, hence to overcome this limitation, we propose a pool-based active learning framework using a fast semi-supervised classifier model termed as Fast Laplacian Twin Support Vector Machine (F Lap − TW SV M) which identifies most informative examples to train the learning model. F Lap − TW SV M is faster than existing Laplacian twin support vector machine as it solves a smaller sized Quadratic Programming Problem (QPP) along with an Unconstrained Minimization Problem (UMP) to obtain decision hyperplanes which can also handle heteroscedastic noise present in the training data. Moreover, the aforementioned framework has been extended to deal with multi-category classification scenarios. Several experiments performed on standard machine learning benchmark datasets along with activity recognition and CBIR systems showed that using only most relevant and informative labeled data can improve the training complexity as well as label requirements when compared to traditional semi-supervised and active learning approaches.

Tensor Learning

Following our approach to use limited labeled data information, we next dealt with activity recognition problem in an unsupervised manner. As the information in activity video sequences is distributed spatially in the form of two-dimensional matrices (elements of second-order tensor space), traditional vector-based approaches rely on low-dimensional features representation for identifying patterns and are thus prone to loss of useful information which is present in the spatial structure of the data. Hence, we propose a novel clustering framework, termed as Ternary Treebased Structural Least Squares Support Tensor Clustering (TT-SLSTWSTC), that builds a cluster model as a hierarchical ternary tree, where at each node nonambiguous data is dealt separately from ambiguous data points using the proposed Ternary Structural Least Squares Support Tensor Machine (TS-LSTWSTM). The TS-LSTWSTM classifier considers the structural risk minimization of data alongside a symmetrical L2-norm loss function. The proposed clustering model helps in identifying relevant patterns in matrix data as they take advantage of structural information present in multidimensional framework and reduce computational overheads as well. Also, initialization framework based on tensor kmeans is used in order to overcome the instability disseminated by random initialization. Computational experiments performed on human activity recognition and handwritten digit recognition problems depicted the comparable performance of this approach to supervised learning