Biologically Inspired Learning Systems
At the Vector Institute's Social AI Lab, I implemented biologically inspired learning systems based on Hebbian and BCM-style plasticity, predictive coding, and sparse representations. The core question was whether useful features could self-organize under local learning rules without explicit backpropagation.
I built custom PyTorch modules for Locally Competitive Algorithm convolutional layers with lateral inhibition, reconstruction loss, and activity regularization. I also developed tooling to inspect layer-wise activation dynamics, measure co-activation statistics, and benchmark stability under different competitive and inhibitory constraints.
The work translated neuroscience-inspired learning rules into runnable experiments, with emphasis on interpretability, self-supervision, and computational mechanisms relevant to energy-efficient neuromorphic systems.