Connecting Computation in Recurrent Neural Networks to Large-scale Spiking Activity

Abstract: New technologies now allow recording the spikes of thousands of neurons during cognitive tasks, providing an unprecedented view of how the brain computes in real time — but analytic tools to fully leverage these datasets are still being developed. During my PhD, I introduced a new decomposition operator for large-scale spike recordings that maps each population spike pattern to a unique vector capturing both the occurrence and precise timing of each spike. In parallel, I developed mathematically solvable recurrent neural networks that perform modern machine learning tasks while remaining fully interpretable. By integrating these approaches, I aim to build a framework for analyzing and predicting neural activity during behavior — one that enables decoding cognitive processes directly from spikes while understanding exactly how these predictions are made.