① Classical bias–variance
— polynomial fit (1-D), the regime you know
truth
avg fit
per-dataset
② Double descent
closed-form min-norm
neural net + GD
— push capacity past p = n (needs high dimensions)
test
bias²
variance
train
‖weights‖₂ — the solution size the min-norm (ℓ₂) bias keeps small
training points n
20
noise σ
0.25
① polynomial degree
3
②
capacity p
20
② ridge λ
0 (off)
② GD iterations T
1000
log-y
Reset