Grokking as the Transition from Lazy to Rich Training Dynamics

TL;DR

This paper explains the grokking phenomenon as a transition from lazy, kernel-like training to rich feature learning in neural networks, highlighting how delayed generalization occurs due to initial misalignment and dataset size.

Contribution

It introduces a new mechanism for grokking based on the transition from lazy to feature learning, supported by analysis of a polynomial regression model and experiments on various architectures.

Findings

Grokking results from a shift from kernel to feature learning.

Delayed generalization depends on initial feature alignment and dataset size.

Transition from lazy to rich training controls grokking in neural networks.

Abstract

We propose that the grokking phenomenon, where the train loss of a neural network decreases much earlier than its test loss, can arise due to a neural network transitioning from lazy training dynamics to a rich, feature learning regime. To illustrate this mechanism, we study the simple setting of vanilla gradient descent on a polynomial regression problem with a two layer neural network which exhibits grokking without regularization in a way that cannot be explained by existing theories. We identify sufficient statistics for the test loss of such a network, and tracking these over training reveals that grokking arises in this setting when the network first attempts to fit a kernel regression solution with its initial features, followed by late-time feature learning where a generalizing solution is identified after train loss is already low. We find that the key determinants of grokking are the rate of feature learning -- which can be controlled precisely by parameters that scale the network output -- and the alignment of the initial features with the target function $y(x)$. We argue this delayed generalization arises when (1) the top eigenvectors of the initial neural tangent kernel and the task labels $y(x)$ are misaligned, but (2) the dataset size is large enough so that it is possible for the network to generalize eventually, but not so large that train loss perfectly tracks test loss at all epochs, and (3) the network begins training in the lazy regime so does not learn features immediately. We conclude with evidence that this transition from lazy (linear model) to rich training (feature learning) can control grokking in more general settings, like on MNIST, one-layer Transformers, and student-teacher networks.