Grokking: When a Neural Network Memorizes First and Understands Later
中文摘要
Grokking是神经网络先通过过拟合进行记忆、随后实现泛化的现象。
English Summary
Grokking is a phenomenon where neural networks first overfit via memorization, then later achieve generalization.
Original Excerpt
A strange deep-learning phenomenon where generalization can appear long after overfitting Continue reading on Medium »