Signal processing & machine learningManas Patil
- M.Tech, Communication Engineering
- Department of Electrical Engineering, IIT Bombay
- Advised by Prof. Vikram M. Gadre
I work on how generative models can be measured — what structure a model is obliged to preserve, and how to detect the structure it quietly loses.
Current work
Most evaluations of a generative model collapse an entire distribution into one number in some embedding space. My work asks whether representations with explicit multiresolution structure — wavelet and framelet decompositions — describe that distribution more stably than embedding-based scores such as FID, FD-DINOv2 and CMMD do. The SPCOM 2026 paper below compares the topology these metrics induce across a large sweep of image generators.
My master's thesis carries this from measurement toward structure: characterising generative models by how they behave under canonical group actions — translation, rotation, dilation, composition — using harmonic analysis, coorbit theory and measure theory on locally compact groups.
Alongside research, I maintain the web and server infrastructure for the Department of Electrical Engineering at IIT Bombay.
Selected publications
Are Wavelets More Reliable? Comparing the Topology of Image Generator Metrics
International Conference on Signal Processing and Communications (SPCOM), IEEE Xplore, 2026Multi-ID Zero Knowledge Proof Systems for Anonymous and Verified Complaints
International Conference for Women in Innovation, Technology and Entrepreneurship (ICWITE), IEEE Xplore, 2025 · doi:10.1109/ICWITE64848.2025.11306926Doc Vault — A Blockchain and Lattice-Cryptography Based Secure Document Storage Platform
International Conference on Emerging Trends in Networks and Computer Communications (ETNCC), IEEE Xplore, 2024 · doi:10.1109/ETNCC63262.2024.10767517
Recent writing
Humanities, Where Are You?
Student politics through a lens of query
On the New Year
A call to myself
Unsolicited Lessons
A diary post about IITB