Deep Learning, Neuromorphic Computing & MicroArchitecture · IIT Madras

Efficient, Robust &
Brain-Inspired AI

BrainSeek Lab works across the full stack of modern deep learning and computer architecture — from microarchitecture security and RISC-V based SoC design to adversarial robustness, model compression, spiking neural networks, and neuromorphic hardware.

View Research Meet the Team

4+ Publications

13+ GPUs across 3 HPC Clusters

16 Team Members

Focus Areas

What We Work On

Our research spans hardware and software — secure microarchitecture and RISC-V SoC design, algorithmic efficiency, robustness, brain-inspired computing, and efficient language models.

🔒

Microarchitecture Security

Investigating side-channel and speculative-execution vulnerabilities at the microarchitecture level, and designing hardware-level mitigations that preserve performance while closing security gaps.

⚙️

RISC-V Based SoC Design

Building custom RISC-V based systems-on-chip tailored for efficient, secure, and neuromorphic-aware compute — bridging our architecture and hardware accelerator research with open ISA design.

🛡️

Adversarial Attacks & Defence

Studying how deep learning models fail under adversarial perturbations and building robust defences — from certified defences to empirical hardening of vision and language models.

📦

Model Compression & Quantization

Making neural networks smaller and faster without sacrificing accuracy. We work on pruning, quantization-aware training, and hardware-aware compression for efficient deployment.

🎓

Knowledge Distillation

Transferring the knowledge of large teacher models into compact student networks — enabling capable models to run on resource-constrained devices.

🗣️

Small Language Models

Designing efficient SLMs (including ShisuLM) for constrained environments. We explore architecture choices, training strategies, and distillation techniques for capable small-scale language models.

⚡

Spiking Neural Networks

Designing and training SNNs for energy-efficient inference — including ANN-SNN conversion (PASCAL), direct training, and biologically plausible learning rules like Equilibrium Propagation.

🔬

Neuromorphic Hardware

Building DL accelerators tailored for sparse, spike-based computation. We exploit dual sparsity (APex) and design quantization schemes for real-time neuromorphic deployment.

🌀

Energy Based Models

Exploring energy-based learning frameworks as principled alternatives to standard supervised training — including Equilibrium Propagation as a biologically plausible credit assignment mechanism.

Latest Work