Deep Networks Always Grok and Here is Why

• URL: http://arxiv.org/abs/2402.15555v2
• Date: Thu, 6 Jun 2024 18:33:15 GMT
• Title: Deep Networks Always Grok and Here is Why
• Authors: Ahmed Imtiaz Humayun, Randall Balestriero, Richard Baraniuk,
• Abstract summary: Grokking, or delayed generalization, is a phenomenon where generalization in a deep neural network (DNN) occurs long after achieving near zero training error. We demonstrate that grokking is actually much more widespread and materializes in a wide range of practical settings.
• Score: 15.327649172531606
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Grokking, or delayed generalization, is a phenomenon where generalization in a deep neural network (DNN) occurs long after achieving near zero training error. Previous studies have reported the occurrence of grokking in specific controlled settings, such as DNNs initialized with large-norm parameters or transformers trained on algorithmic datasets. We demonstrate that grokking is actually much more widespread and materializes in a wide range of practical settings, such as training of a convolutional neural network (CNN) on CIFAR10 or a Resnet on Imagenette. We introduce the new concept of delayed robustness, whereby a DNN groks adversarial examples and becomes robust, long after interpolation and/or generalization. We develop an analytical explanation for the emergence of both delayed generalization and delayed robustness based on the local complexity of a DNN's input-output mapping. Our local complexity measures the density of so-called linear regions (aka, spline partition regions) that tile the DNN input space and serves as a utile progress measure for training. We provide the first evidence that, for classification problems, the linear regions undergo a phase transition during training whereafter they migrate away from the training samples (making the DNN mapping smoother there) and towards the decision boundary (making the DNN mapping less smooth there). Grokking occurs post phase transition as a robust partition of the input space thanks to the linearization of the DNN mapping around the training points. Website: https://bit.ly/grok-adversarial

Related papers

• On Generalization Bounds for Deep Compound Gaussian Neural Networks [1.4425878137951238]
  Unrolled deep neural networks (DNNs) provide better interpretability and superior empirical performance than standard DNNs. We develop novel generalization error bounds for a class of unrolled DNNs informed by a compound Gaussian prior. Under realistic conditions, we show that, at worst, the generalization error scales $mathcalO(nsqrt(n))$ in the signal dimension and $mathcalO(($Network Size$)3/2)$ in network size.
  arXiv  Detail & Related papers  (2024-02-20T16:01:39Z)
• DNN Training Acceleration via Exploring GPGPU Friendly Sparsity [16.406482603838157]
  We propose the Approximate Random Dropout that replaces the conventional random dropout of neurons and synapses with a regular and online generated row-based or tile-based dropout patterns. We then develop a SGD-based Search Algorithm that produces the distribution of row-based or tile-based dropout patterns to compensate for the potential accuracy loss. We also propose the sensitivity-aware dropout method to dynamically drop the input feature maps based on their sensitivity so as to achieve greater forward and backward training acceleration.
  arXiv  Detail & Related papers  (2022-03-11T01:32:03Z)
• Disentangling deep neural networks with rectified linear units using duality [4.683806391173103]
  We propose a novel interpretable counterpart of deep neural networks (DNNs) with rectified linear units (ReLUs) We show that convolution with global pooling and skip connection provide respectively rotational invariance and ensemble structure to the neural path kernel (NPK)
  arXiv  Detail & Related papers  (2021-10-06T16:51:59Z)
• Generalizing Neural Networks by Reflecting Deviating Data in Production [15.498447555957773]
  We present a runtime approach that mitigates DNN mis-predictions caused by unexpected runtime inputs to the DNN. We use a distribution analyzer based on the distance metric learned by a Siamese network to identify "unseen" semantically-preserving inputs. Our approach transforms those unexpected inputs into inputs from the training set that are identified as having similar semantics.
  arXiv  Detail & Related papers  (2021-10-06T13:05:45Z)
• Shift-Robust GNNs: Overcoming the Limitations of Localized Graph Training data [52.771780951404565]
  Shift-Robust GNN (SR-GNN) is designed to account for distributional differences between biased training data and the graph's true inference distribution. We show that SR-GNN outperforms other GNN baselines by accuracy, eliminating at least (40%) of the negative effects introduced by biased training data.
  arXiv  Detail & Related papers  (2021-08-02T18:00:38Z)
• Wide and Deep Graph Neural Network with Distributed Online Learning [174.8221510182559]
  Graph neural networks (GNNs) are naturally distributed architectures for learning representations from network data. Online learning can be leveraged to retrain GNNs at testing time to overcome this issue. This paper develops the Wide and Deep GNN (WD-GNN), a novel architecture that can be updated with distributed online learning mechanisms.
  arXiv  Detail & Related papers  (2021-07-19T23:56:48Z)
• Understanding and Improving Early Stopping for Learning with Noisy Labels [63.0730063791198]
  The memorization effect of deep neural network (DNN) plays a pivotal role in many state-of-the-art label-noise learning methods. Current methods generally decide the early stopping point by considering a DNN as a whole. We propose to separate a DNN into different parts and progressively train them to address this problem.
  arXiv  Detail & Related papers  (2021-06-30T07:18:00Z)
• Redundant representations help generalization in wide neural networks [71.38860635025907]
  We study the last hidden layer representations of various state-of-the-art convolutional neural networks. We find that if the last hidden representation is wide enough, its neurons tend to split into groups that carry identical information, and differ from each other only by statistically independent noise.
  arXiv  Detail & Related papers  (2021-06-07T10:18:54Z)
• Online Limited Memory Neural-Linear Bandits with Likelihood Matching [53.18698496031658]
  We study neural-linear bandits for solving problems where both exploration and representation learning play an important role. We propose a likelihood matching algorithm that is resilient to catastrophic forgetting and is completely online.
  arXiv  Detail & Related papers  (2021-02-07T14:19:07Z)
• Overcoming Catastrophic Forgetting in Graph Neural Networks [50.900153089330175]
  Catastrophic forgetting refers to the tendency that a neural network "forgets" the previous learned knowledge upon learning new tasks. We propose a novel scheme dedicated to overcoming this problem and hence strengthen continual learning in graph neural networks (GNNs) At the heart of our approach is a generic module, termed as topology-aware weight preserving(TWP)
  arXiv  Detail & Related papers  (2020-12-10T22:30:25Z)

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.