(Multiresolution) Signal Processing for Interpretability and Economy in Generative AI

Basic Information

Project Title / Domain: (Multiresolution) Signal Processing for Interpretability and Economy in Generative AI
Name: Abhijat Bharadwaj
Guide: Prof. Vikram M. Gadre
Project Type: Both. Part theory, part experimentation.

Short Description of Project

The project revolves around making the AI/ML/DL systems (1) more interpretable and (2) more economic and efficient. The problem of interpretability/explainability can range from trying to explain the behaviour of already built models (post-hoc) or designing models or frameworks which are inherently interpretable (ante-hoc). The problem of economy is mostly associated with reducing the number of trainable parameters for a task while maintaining the model performance. My work has revolved mostly around image generation models. The research focuses of exploring signal processing inspired ideas to achieve the aforementioned goals.

Whom did you work with?

PG students, The prof directly, Peers (Other UG students)

Tools / Simulation / Software / Hardware

Standard ML and software tools.

Expectations from Guide

I need weekly meetings and theoretical discussions on my approach to validate them.

Expectations for 8th Sem & Summer

In the 8th semester, it was expected that I catch up on the literature survey and prior works, and decide or atleast steer towards a concrete problem statement with some evidence to support my hypothesis.

I had a summer intern, with clearance from my guide, and did not stay in insti.

Load: 8th Sem vs 9th (Placement) Sem

I may not be the best guy to answer this. I did not sit for placements and was able to spend more time on my DDP in the 9th sem.

However, in general, the professor is understanding of the placement scenario. The expectations were set higher during the beginning of the 9th sem, much more than that of the 8th sem, and the workload was lowered significantly as the placements approached.

Summer on Campus

Is DDP Guide Same as SRE/RnD Guide?

Reason for Change of Guide

Liked this project better.

How did your SRE help with your DDP?

End Deliverables

I’m focusing more on the interpretability side of things so the deliverables are open-ended, mostly on finding ways in which (multiresolution) signal processing techniques can assist the development of AI/ML systems.

Advice

Maintain a clear and open communication with your guide. It helps balancing expectations from both the sides.