Projects

A comprehensive spiking neural network (SNN) framework built on biologically realistic leaky integrate-and-fire neurons. Features spike-timing-dependent plasticity (STDP), surrogate gradient training, and benchmarks against conventional deep learning on temporal pattern recognition tasks.

Neural architecture search powered by CMA-ES (Covariance Matrix Adaptation Evolution Strategy) and multi-objective genetic programming. Discovers efficient transformer and CNN variants without gradient-based search, inspired by natural selection principles.

A hardware-aware, reproducible evaluation harness for large language models. Covers 15+ benchmarks (MMLU, HumanEval, GSM8K, ARC, BIG-Bench Hard), with automated reporting, calibration analysis, and efficiency metrics.

Neural memory architecture inspired by the hippocampal-entorhinal circuit. Implements pattern completion, pattern separation, and episodic memory replay to achieve continual learning without catastrophic forgetting.

Swarm intelligence library for hyperparameter optimisation and neural network training. Implements PSO, ACO, and bee-algorithm variants, demonstrating competitive performance with gradient-free training on non-differentiable loss landscapes.

Interpretability toolkit for transformer language models. Implements attention visualisation, activation patching, logit lens, and probing classifiers to understand what LLMs learn and how they represent knowledge.