S. Sumitra

Silpa Vadakkeeveetil Sreelatha, Adarsh Kappiyath, S Sumitra

Several methods for discovering interpretable directions in the latent space of pre-trained GANs have been proposed. Latent semantics discovered by unsupervised methods are relatively less disentangled than supervised methods since they do not use pre-trained attribute classifiers. We propose Scale Ranking Estimator (SRE), which is trained using self-supervision. SRE enhances the disentanglement in directions obtained by existing unsupervised disentanglement techniques. These directions are updated to preserve the ordering of variation within each direction in latent space. Qualitative and quantitative evaluation of the discovered directions demonstrates that our proposed method significantly improves disentanglement in various datasets. We also show that the learned SRE can be used to perform Attribute-based image retrieval task without further training.

Akhil Meethal, Asharaf S, Sumitra S

Kernel based Deep Learning using multi-layer kernel machines(MKMs) was proposed by Y.Cho and L.K. Saul in \cite{saul}. In MKMs they used only one kernel(arc-cosine kernel) at a layer for the kernel PCA-based feature extraction. We propose to use multiple kernels in each layer by taking a convex combination of many kernels following an unsupervised learning strategy. Empirical study is conducted on \textit{mnist-back-rand}, \textit{mnist-back-image} and \textit{mnist-rot-back-image} datasets generated by adding random noise in the image background of MNIST dataset. Experimental results indicate that using MKL in MKMs earns a better representation of the raw data and improves the classifier performance.

CR Sandeep, Asif Salim, R Sethunadh et al.

The process of selecting points for training a machine learning model is often a challenging task. Many times, we will have a lot of data, but for training, we require the labels and labeling is often costly. So we need to select the points for training in an efficient manner so that the model trained on the points selected will be better than the ones trained on any other training set. We propose a novel method to select the nodes in graph datasets using the concept of graph centrality. Two methods are proposed - one using a smart selection strategy, where the model is required to be trained only once and another using active learning method. We have tested this idea on three popular graph datasets - Cora, Citeseer and Pubmed- and the results are found to be encouraging.

Asif Salim, Shiju. S. S, Sumitra. S

We describe the design of a reproducing kernel suitable for attributed graphs, in which the similarity between the two graphs is defined based on the neighborhood information of the graph nodes with the aid of a product graph formulation. We represent the proposed kernel as the weighted sum of two other kernels of which one is an R-convolution kernel that processes the attribute information of the graph and the other is an optimal assignment kernel that processes label information. They are formulated in such a way that the edges processed as part of the kernel computation have the same neighborhood properties and hence the kernel proposed makes a well-defined correspondence between regions processed in graphs. These concepts are also extended to the case of the shortest paths. We identified the state-of-the-art kernels that can be mapped to such a neighborhood preserving framework. We found that the kernel value of the argument graphs in each iteration of the Weisfeiler-Lehman color refinement algorithm can be obtained recursively from the product graph formulated in our method. By incorporating the proposed kernel on support vector machines we analyzed the real-world data sets and it has shown superior performance in comparison with that of the other state-of-the-art graph kernels.

Asif Salim, Sumitra S

Graph convolutional neural networks (GCNNs) have been widely used in graph learning. It has been observed that the smoothness functional on graphs can be defined in terms of the graph Laplacian. This fact points out in the direction of using Laplacian in deriving regularization operators on graphs and its consequent use with spectral GCNN filter designs. In this work, we explore the regularization properties of graph Laplacian and proposed a generalized framework for regularized filter designs in spectral GCNNs. We found that the filters used in many state-of-the-art GCNNs can be derived as a special case of the framework we developed. We designed new filters that are associated with well-defined regularization behavior and tested their performance on semi-supervised node classification tasks. Their performance was found to be superior to that of the other state-of-the-art techniques.

Silpa Vadakkeeveetil Sreelatha, Adarsh Kappiyath, Sumitra S

We propose Disentanglement based Active Learning (DAL), a new active learning technique based on self-supervision which leverages the concept of disentanglement. Instead of requesting labels from human oracle, our method automatically labels the majority of the datapoints, thus drastically reducing the human labeling budget in Generative Adversarial Net (GAN) based active learning approaches. The proposed method uses Information Maximizing Generative Adversarial Nets (InfoGAN) to learn disentangled class category representations. Disagreement between active learner predictions and InfoGAN labels decides if the datapoints need to be human-labeled. We also introduce a label correction mechanism that aims to filter out label noise that occurs due to automatic labeling. Results on three benchmark datasets for the image classification task demonstrate that our method achieves better performance compared to existing GAN-based active learning approaches.

Rahul Vashisht, H. Viji, T. Sundararajan et al.

The advancement of machine learning algorithms has opened a wide scope for vibration-based SHM (Structural Health Monitoring). Vibration-based SHM is based on the fact that damage will alter the dynamic properties viz., structural response, frequencies, mode shapes, etc of the structure. The responses measured using sensors, which are high dimensional in nature, can be intelligently analyzed using machine learning techniques for damage assessment. Neural networks employing multilayer architectures are expressive models capable of capturing complex relationships between input-output pairs but do not account for uncertainty in network outputs. A BNN (Bayesian Neural Network) refers to extending standard networks with posterior inference. It is a neural network with a prior distribution on its weights. Deep learning architectures like CNN (Convolutional neural network) and LSTM(Long Short Term Memory) are good candidates for representation learning from high dimensional data. The advantage of using CNN over multi-layer neural networks is that they are good feature extractors as well as classifiers, which eliminates the need for generating hand-engineered features. LSTM networks are mainly used for sequence modeling. This paper presents both a Bayesian multi-layer perceptron and deep learning-based approach for damage detection and location identification in beam-like structures. Raw frequency response data simulated using finite element analysis is fed as the input of the network. As part of this, frequency response was generated for a series of simulations in the cantilever beam involving different damage scenarios. This case study shows the effectiveness of the above approaches to predict bending rigidity with an acceptable error rate.