Grokking as the Transition from Lazy to Rich Training Dynamics

• URL: http://arxiv.org/abs/2310.06110v3
• Date: Thu, 11 Apr 2024 16:15:34 GMT
• Title: Grokking as the Transition from Lazy to Rich Training Dynamics
• Authors: Tanishq Kumar, Blake Bordelon, Samuel J. Gershman, Cengiz Pehlevan,
• Abstract summary: grokking occurs when the train loss of a neural network decreases much earlier than its test loss. Key determinants of grokking are the rate of feature learning and the alignment of the initial features with the target function.
• Score: 35.186196991224286
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• Benign Overfitting for Two-layer ReLU Convolutional Neural Networks [60.19739010031304]
  We establish algorithm-dependent risk bounds for learning two-layer ReLU convolutional neural networks with label-flipping noise. We show that, under mild conditions, the neural network trained by gradient descent can achieve near-zero training loss and Bayes optimal test risk.
  arXiv  Detail & Related papers  (2023-03-07T18:59:38Z)
• Theoretical Characterization of How Neural Network Pruning Affects its Generalization [131.1347309639727]
  This work makes the first attempt to study how different pruning fractions affect the model's gradient descent dynamics and generalization. It is shown that as long as the pruning fraction is below a certain threshold, gradient descent can drive the training loss toward zero. More surprisingly, the generalization bound gets better as the pruning fraction gets larger.
  arXiv  Detail & Related papers  (2023-01-01T03:10:45Z)
• Neural networks trained with SGD learn distributions of increasing complexity [78.30235086565388]
  We show that neural networks trained using gradient descent initially classify their inputs using lower-order input statistics. We then exploit higher-order statistics only later during training. We discuss the relation of DSB to other simplicity biases and consider its implications for the principle of universality in learning.
  arXiv  Detail & Related papers  (2022-11-21T15:27:22Z)
• Improving the Trainability of Deep Neural Networks through Layerwise Batch-Entropy Regularization [1.3999481573773072]
  We introduce and evaluate the batch-entropy which quantifies the flow of information through each layer of a neural network. We show that we can train a "vanilla" fully connected network and convolutional neural network with 500 layers by simply adding the batch-entropy regularization term to the loss function.
  arXiv  Detail & Related papers  (2022-08-01T20:31:58Z)
• Benign Overfitting in Two-layer Convolutional Neural Networks [90.75603889605043]
  We study the benign overfitting phenomenon in training a two-layer convolutional neural network (CNN) We show that when the signal-to-noise ratio satisfies a certain condition, a two-layer CNN trained by gradient descent can achieve arbitrarily small training and test loss. On the other hand, when this condition does not hold, overfitting becomes harmful and the obtained CNN can only achieve constant level test loss.
  arXiv  Detail & Related papers  (2022-02-14T07:45:51Z)
• Why Lottery Ticket Wins? A Theoretical Perspective of Sample Complexity on Pruned Neural Networks [79.74580058178594]
  We analyze the performance of training a pruned neural network by analyzing the geometric structure of the objective function. We show that the convex region near a desirable model with guaranteed generalization enlarges as the neural network model is pruned.
  arXiv  Detail & Related papers  (2021-10-12T01:11:07Z)
• Over-parametrized neural networks as under-determined linear systems [31.69089186688224]
  We show that it is unsurprising simple neural networks can achieve zero training loss. We show that kernels typically associated with the ReLU activation function have fundamental flaws. We propose new activation functions that avoid the pitfalls of ReLU in that they admit zero training loss solutions for any set of distinct data points.
  arXiv  Detail & Related papers  (2020-10-29T21:43:00Z)
• It's Hard for Neural Networks To Learn the Game of Life [4.061135251278187]
  Recent findings suggest that neural networks rely on lucky random initial weights of "lottery tickets" that converge quickly to a solution. We examine small convolutional networks that are trained to predict n steps of the two-dimensional cellular automaton Conway's Game of Life. We find that networks of this architecture trained on this task rarely converge.
  arXiv  Detail & Related papers  (2020-09-03T00:47:08Z)
• A Revision of Neural Tangent Kernel-based Approaches for Neural Networks [34.75076385561115]
  We use the neural tangent kernel to show that networks can fit any finite training sample perfectly. A simple and analytic kernel function was derived as indeed equivalent to a fully-trained network. Our tighter analysis resolves the scaling problem and enables the validation of the original NTK-based results.
  arXiv  Detail & Related papers  (2020-07-02T05:07:55Z)
• The Surprising Simplicity of the Early-Time Learning Dynamics of Neural Networks [43.860358308049044]
  In work, we show that these common perceptions can be completely false in the early phase of learning. We argue that this surprising simplicity can persist in networks with more layers with convolutional architecture.
  arXiv  Detail & Related papers  (2020-06-25T17:42:49Z)
• Feature Purification: How Adversarial Training Performs Robust Deep Learning [66.05472746340142]
  We show a principle that we call Feature Purification, where we show one of the causes of the existence of adversarial examples is the accumulation of certain small dense mixtures in the hidden weights during the training process of a neural network. We present both experiments on the CIFAR-10 dataset to illustrate this principle, and a theoretical result proving that for certain natural classification tasks, training a two-layer neural network with ReLU activation using randomly gradient descent indeed this principle.
  arXiv  Detail & Related papers  (2020-05-20T16:56:08Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.