vidhyakamakshi

NITC is a vibrant place that offers best oppurtunites for researchers and I being a member of this community would resonate the same. My research interests are specifically in the field of Machine, Learning, Deep Learning, Computer Vision, etc. My PhD thesis centers around explainability for computer vision models whose summary is presented below for your quick reference:

Image classifcation is the task in which a machine learning model predicts the class/category of an object contained in a given image from the set of known classes. Convolutional Neural Networks(CNN) have achieved state-of-the-art image classifcation results. However, how these models arrive at predictions for a given image is unclear.

The research field Explainable AI (XAI), aims to unravel the working mechanism used by these accurate, opaque black boxes. If the explanations are closer to how humans interpret images, they help better understand the working mechanism of CNNs. This fact was proved by previous experiments as reviewed from existing XAI literature. Studies show that humans process images in terms of sub-regions called concepts. For instance, a peacock is identifed by its characteristic concepts like green feathers, blue neck, etc.

My thesis intended to automatically extract such concepts learned by CNN from the data. Three novel frameworks are proposed to provide automatically extracted concept-based explanations for standard image classifers. The first framework, PACE, automatically extracts class-specifc concepts relevant to the prediction. While class-specifc concepts unravel the blueprints of a class from CNN’s perspective, concepts are often shared across classes; for instance, gorillas and chimpanzees naturally share many characteristics as they belong to the same family. The second framework, SCE, unravels the concept sharedness across related classes from CNNs perspective. The relevance of the extracted concepts towards prediction and the primitive image aspects, like color, texture, and shape encoded by the concept, are estimated after training the explainer.

The thesis identifes a void in XAI’s panorama that much attention is given to classifers trained and tested using the same data. However, allied paradigms have been shown to add to state-of-the-art successes. Despite the data hunger of deep models, domain adaptation techniques have been employed to leverage a huge amount of related data to help learn a classifer that is expected to work on scarce data of interest. The third framework XSDA-Net, builds a supervised domain-adapted classifer that can explain itself in terms of concepts extracted from the diferent datasets the classifer is exposed to.

Experiments demonstrate the utility of all three proposed frameworks in automatically extracting concepts from the data such that they unravel the working mechanism of the image classifers. The thesis reviews the diferent types of explanations prevalent in the XAI feld and enlightens the possible future research avenues for potential researchers looking to venture into XAI.

Given that I work in explainability which is crucial for any state of the art deep learning system and have started branching out into allied learning paradigms by proposing mechanisms that help harness the fruits of explainability and the paradigm, there is a possibility to branch out into any deep learning paradigms and build models that are explainable, fair, transparent and accountable which are necessary characteristics a model is expected to exhibit for deployment in real-world safety-critical environments.