1 paper · 1 filter
Elliot Paquette, Ke Liang Xiao, Yizhe Zhu
Scaling laws for neural networks, in which the loss decays as a power-law in the number of parameters, data, and compute, depend fundamentally on the spectral structure of the data…