Mathematics Colloquia and Seminars

Return to Colloquia & Seminar listing

Synaptic pruning has been explored in both biological and artificial settings, but with different goals and constraints. For example, most pruning algorithms prioritize faster computation and/or reduced storage requirements while maintaining performance in a task-dependent setting, eschewing biological plausibility. As such, these pruning algorithms are unlikely to elucidate the brain’s pruning mechanisms. In this talk, I will present biologically plausible pruning rules that facilitate efficient artificial neural network computation. By utilizing the matrix Chernoff inequality, I will show that a noise-driven pruning approach approximately preserves the spectrum of a limited class of recurrent neural networks while drastically reducing parameter storage requirements. Afterward, I will present another stochastic pruning algorithm using the rectangular matrix Bernstein inequality, providing similar theoretical guarantees for a wider class of neural networks. My results indicate that these recently-developed random matrix inequalities should be of special interest to neuroscientists and machine learning scientists alike, potentially strengthening the bond between these fields.

GGAM Exit Seminar