Search Papers | Poster Sessions | All Posters
Poster C96 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
A Minimal and Flexible Model for Investigating Critical Learning Periods
Sebastian Lee1,2, Stefano Sarao Mannelli2, Andrew Saxe2; 1Imperial College London, 2University College London
We present a mathematically solvable model of critical learning periods using the teacher-student setup from statistical physics. By separating training into `disrupted' and `true' learning regimes, we model several possible learning perturbations. Preliminary results in this model provide evidence of critical learning periods resembling those observed across neuroscience and deep learning, thereby laying the foundations for a theory of critical learning periods.
Keywords: critical learning periods deep learning theory statistical physics