Research

My broad research areas are Deep Learning and Computer Vision. My research interests are are chiefly related to the robustness, adaptability, and interpret-ability of the Deep Neural Networks (DNNs).

1. Data-Free Knowledge Transfer/Distillation (Adaptability)

Knowledge distillation deals with the problem of training a smaller model (Student) from a high capacity source model (Teacher) so as to retain most of its performance. Existing approaches use either the training data or meta-data extracted from it in order to train the Student. However, accessing the dataset on which the Teacher has been trained may not always be feasible if the dataset is very large or it poses privacy or safety concerns (e.g., bio-metric or medical data). Hence, in this project, we aim to develop novel data-free methods to train the Student from the Teacher. Without even using any meta-data about the target dataset, we attempt to synthesise the synthetic samples (e.g. Data Impressions) from the complex Teacher model and utilise these as surrogates for the original training data samples to transfer its learning to Student via knowledge distillation. Therefore, we dub this procedure "Zero-Shot Knowledge Distillation".

2. Robustness of Deep Visual representations (Robustness)

Machine learning systems are vulnerable to adversarial samples - malicious input with structured perturbations that can fool the systems to infer wrong predictions. Recently, Deep Convolutional Neural Network (CNN) based object classifiers are also shown to be fooled by adversarial perturbations that are quasi-imperceptible to humans. There have been multiple approaches formulated to compute the adversarial samples, exploiting linearity of the models, finite training data, etc. More importantly, adversarial samples can be transferred (generalized) from one model to another, even if the second model has a different architecture and trained on different subset of training data. This property allows an attacker to launch an attack without the knowledge of the target model’s internals, which makes them a dangerous threat for deploying the models in practice. Particularly, for critical applications that involve safety, robust models should be learned towards adversarial attacks. Therefore, the effect of adversarial perturbations warrants the need for in depth analysis of this subject.

Deep learning Models are complex machine learning systems with hundreds of layers and millions of parameters. Presence of advanced regularizers such as dropout and batch-normalization make the models less transparent. Because of end-to-end nature of the learning, models suffer from lesser decomposability and hence most of us treat them as black-boxes. In order to make these models more transparent, we devise methods that provide visual explanations to the labels predicted by the recognition CNNs.

CNN Fixations

(Trans. on Image Processing 2019)

4. Incremental Learning (Adaptability)

Deep Neural Networks suffer from catastrophic forgetting, dramatic fall in the performance on the old categories when new classes are added incrementally. This can pose serious challenges affecting their real-world deployment where new classes are added incrementally. An easy example is a commercial face recognition system, where it has to learn to recognise new faces without having to forget the old faces. The existing training setup that requires all the training samples (old+new) to update the model is not sustainable as the number of classes grows. Hence we attempt to device methods to efficiently handle Incremental Learning and to enable the models towards continual learning.

5. Augmenting Deep Visual Representations from side and additional information (Adaptability)

Deep Convolutional neural networks (CNNs) have resulted in unprecedented performances for visual recognition. They have been shown to learn representations that can efficiently discriminate hundreds of visual categories. In their vanilla supervised setting, CNNs learn from large scale datasets that offer category labels. I want to exploit the useful “side and additional information” to enrich the representations with more semantics. Specifically, I have been devising approaches to encode additional discriminative information from cues such as (i) objectness, (ii) textual tags associated with images and (iii) strong supervision offered by the captions.