Projects

Some (Old Undergrad) Research Projects



We obtain a better convergence rate

Accelerating MCMC

With Prof. Vivek Borkar, IIT Bombay

Random-walk based approaches are useful for estimating the average value of some mapping in undirected graphs but these algorithms tend to get stuck in local clusters. Similar problem is encountered by MCMC algorithms as well.

We proposed a novel modification to algorithms which overcomes the problem of getting stuck in local clusters in highly clustered graphs. The proposed algorithm outperforms other prevalent algorithms significantly on well-clustered graphs.

Explainable AI in Text Classification

With Prof. Gitta Kutyniok, Technische Universität Berlin

Explainable AI (XAI) provides an insight and tries to explain the decisions of deep neural networks which are otherwise taken to be black-boxes. Using the knowledge of the neural network's decision process, we can focus on how to fix its shortcomings and how to make it focus on certain features that we believe are relevant for the decision process.

Our proposed method was inspired from the previous work by Prof. Kutyniok in XAI using rate-distortion theory. We built on top of the rate distortion based method and proposed a novel developed a novel technique which outperforms current state-of-the-art in XAI. We further explored the scope of using XAI to boost the performance of neural networks fully automatically, and its scope in model compression.



Superior performance of our technique!



Block diagram for the self-attention block

With Prof. Ram Nevatia, University of Southern California

Image forgery detection is an important task with the growing popularity and ease of access to professional tools like Photoshop. The work focuses on detection as well as localization of several types of image manipulations.

Our proposed method is inspired from the revolutionary transformer network introduced in NLP, which is based on self-attention. We extend this technique for images and effectively devise a strategy for positional embedding along with appropriate methods to obtain the Q,K and V values corresponding to the transformer architecture. The proposed technique achieves state-of-the-art.


With Prof. Subhasis Chaudhuri, IIT Bombay

The project deals with the problem of semantic classification of challenging and highly-cluttered dataset. We present a novel, and yet a very simple classification technique by leveraging the ease of classifiability of any existing well separable dataset for guidance. Since the guide dataset which may or may not have any semantic relationship with the experimental dataset, forms well separable clusters in the feature set, the proposed network tries to embed class-wise features of the challenging dataset to those distinct clusters of the guide set, making them more separable. Depending on the availability, we proposed two types of guide sets: one using texture (image) guides and another using prototype vectors representing cluster centers. Experimental results obtained on the challenging datasets establish the efficacy of the proposed method as we could outperform the existing state-of-the-art techniques by a considerable margin.






Clustering cluttered data using a guide



Compound Arbitrarily Varying Channel

With Prof. Bikash Dey, IIT Bombay, and Prof. Vinod Prabhakaran, TIFR

We propose a novel communication model called Compound Arbitrarily Varying Channels (C-AVCs) which is a unifying generalization of compound-DMCs and Arbitrarily Varying Channels (AVCs). The C-AVC can be viewed as a channel model where multiple adversaries are present but only one of them operates during a transmission block. The active adversary can induce a family of channels arbitrarily for each symbol of transmission. 

Other can communication, a new problem of adversary identification also arises for a C-AVC. We have analyzed the C-AVC capacity for three distinct but closely related problems under both random and deterministic coding. This project is a part of my Bachelor's Thesis and it has been submitted to ISIT 2021.

Estimation of Asymptomatic Population from Symptomatic Population

With Prof. Vivek Borkar, IIT Bombay

Accurate prediction of asymptomatic COVID-19 patients can help Government agencies take appropriate recourse. Our aim is to develop a method to predict an unobservable population from an observed time series.

We proposed a two-stage strategy which first fits an underlying Hidden Markov Model and then predicts the unobserved components. We used a 4-compartment, 5-parameter based underlying model. We aim to estimate the unobserved data using a stochastic approximation variant of Expectation-Maximization algorithm.





With Prof. Preeti Rao, IIT Bombay

Syllable detection is an important speech analysis task with applications in speech rate estimation, word segmentation, and automatic prosody detection. We proposed a pipelined signal processing approach to detect the syllables and hence, estimate the speech rate. This pipelined approach is extremely useful when the amount of data available is not feasible to train an end-to-end neural network.

The proposed method uses weighted MFCC features from the input signal to obtain an energy contour on which we perform operations like peak detection on obtain syllable location. The resulting non-convex, non-closed-form cost function was optimized using Particle Swarm Optimization.

Course & Other Projects

Speech Command Classifier

Pipelined RISC Microprocessor Design

Image Deblurring

Electronic Violin

Acoustic Source Locator

Solving Inverse Problems in Computer Vision

Maze Solver Using MDP Planning

Web Crawler