The authors conducted experiments on CIFAR10, using cats and dogs as a clean classification task while 20% of training images (automobiles) were assigned fixed random labels.
Kernel geometry determines whether early stopping beats interpolation under label noise
The authors derive a phase diagram for kernel gradient flow showing regimes where test error is minimized at finite time or only at interpolation, and identify a critical geometry where interpolation is optimal.
Independent
Fabrizio Boncoraglio · Lenka Zdeborová
Research Digest··3 min read
The authors study when early stopping is beneficial under label noise, challenging the common intuition that it is always necessary.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code