Neuromorphic computing · Spiking neural networks
Research & Innovation Engineer at TCS Research, Kolkata
I build spiking neural networks, and the tooling that moves them between neuromorphic frameworks without quietly changing what they compute. At the moment I am trying to find where spiking models fail at language, and whether adaptive neurons are any help.
Trajectory
It started with vision, picked up language along the way, and both ended up asking the same question about where computation should be spent.
Adani Institute of Infrastructure Engineering, Ahmedabad
Four years of signals, networks and circuits before any of it was called machine learning. Graduated at 9.22, cleared GATE at 93.90 percentile, and left with the habit of asking what a system costs before asking what it can do.
IIIT Delhi
Two years of coursework and research in AI, alongside teaching assistantships for NLP, DSA, AI and Intelligent Systems. The lab work split into two directions that only converged much later.
MIDAS Lab, IIIT Delhi · with Dr. Rajiv Ratn Shah
Crowd energy from audio and video, published at IEEE MIPR. Person attribute recognition for Indian street scenes, put on arXiv. Accident detection from live video with TU Darmstadt. All three had the same shape: a model that works offline, and a stream that will not wait for it.
IIIT Delhi · with Dr. Shad Akhtar
Perspective-controlled summarisation of healthcare threads, using classifier guidance and an energy term over adapters and prefix tuning. The interesting part was not the score. It was watching how little of a large model you actually need to move if you are precise about which part you move.
TCS Research, Kolkata · with Dr. Barnali Basak and Sounak Dey
The main project is a bridge between general neuromorphic frameworks and hardware-specific ones: one spiking model, ported across SNNTorch, Rockpool, Lava and Sinabs, holding 96% accuracy within 2% of the reference. Building it meant first building real intuition for how SNNs behave — how event-based data should be handled, and where sparsity actually buys you something in multimodal tasks rather than just looking elegant on a slide.
Current exploration
Language is the hard case for spiking networks, so that is where I am looking for the cracks. Alongside it I am working through adaptive-LIF neurons — how their threshold adaptation changes what a network can hold over time, and whether that is something I can bring back into the porting project rather than leave as a side experiment.
Spiking networks only win if the silence is worth something. I work on event-based data and on finding the multimodal tasks where sparse, event-driven computation is genuinely cheaper, not just biologically flattering.
Text is the awkward case for spiking networks: no natural event stream, long dependencies, dense embeddings. I am mapping where spiking language models actually break, which is more useful right now than another benchmark number.
Adaptive-LIF neurons move their own threshold as they fire. That gives a network a form of memory the weights never see, and I want to know how much of it can be relied on.
SNNTorch, Lava, Rockpool and Sinabs each assume something different about time, reset and quantisation. A result that cannot survive the move between them is a result about a framework.
Full record Google Scholar ORCID 0009-0001-8159-2516
IEEE 7th International Conference on Multimedia Information Processing and Retrieval (MIPR), pp. 308–314 · doi:10.1109/MIPR62202.2024.00056
My partI built the audio half: the CLAP-based classifier for crowd noise, and the 90-minute labelled dataset it trains on.
arXiv:2407.03305
My partI worked on the attribute recognition model and the cascade that made it fast enough to run on a live feed.
A bridge between general neuromorphic frameworks and hardware-specific ones, plus a harness that measures what a port actually costs.
Steering healthcare summaries towards a requested perspective, using classifier guidance and an energy term over adapters and prefix tuning.
An agent that walks a knowledge graph to answer a query. 95.10 AUC-PR on Countries-S3, the hardest split of the three.
A practical beginner-friendly tutorial on building Spiking Neural Networks using SNNTorch.
A short introduction to neuromorphic computing and why it matters for the future of AI.
If you work on neuromorphic hardware and algorithms, efficient deep learning, or spiking models for language and perception, write to me. I am glad to send a longer version of any of this. Corrections are welcome too; a fair amount of what is on this site is still provisional.