double-descent

OpenAI research reveals the 'double descent' phenomenon where test error exhibits a non-monotonic pattern as both model size and training steps increase, challenging traditional understanding of the bias-variance tradeoff in deep learning.