Research

Overview

The Question: How do biological circuits implement the computation required for flexible behavior?

My work combines mathematical theory, high-performance computing, and biological constraints to investigate the mechanistic basis of neural computation. I am specifically interested in how neural circuits infer and maintain latent representations of the environment to support zero-shot and few-shot learning.

Core Themes

1. Neural Manifolds & Mechanistic Substrates

While the field often studies neural manifolds in abstract rate-based models, I aim to anchor these concepts in biological constraints.

• Goal: Develop a bridge between low-dimensional population dynamics and specific connectivity motifs in Spiking Neural Networks (SNNs).
• Approach: Using “in silico” models as testbeds to unify control theory, dynamical systems, and STDP-based learning rules.

2. Latent State Inference & Cognitive Flexibility

Animals can adapt to novel scenarios with minimal experience, suggesting they use “latent states” to represent context.

• Hypothesis: Circuits do not just overwrite synaptic weights during learning; they shift and modulate existing manifolds to represent new task rules.
• Application: Studying Reversal Learning and Flexible Decision-Making to understand how networks balance stability (remembering old tasks) with plasticity (learning new ones) without catastrophic interference.

Toolbox

• Theory: Dynamical Systems, Graph Theory, Information Geometry.
• Computation: JAX/Flax for differentiable physics; CUDA for parallel acceleration.
• Modeling: Leaky Integrate-and-Fire (LIF) networks, Recurrent Neural Networks (RNNs), Evolutionary Optimization.