Partially trainable networks show a sharp approximation phase transition
For spherical biased-ReLU features at d=5 and τ=10, about 4.1% trainable width captures 90% of the bound reduction.
Stories per day · 30 days
1 stories, newest first · one card per story