Do Deep Networks Forget Initialization? A Forgetting-Time View of Practical Inductive Bias

• URL: http://arxiv.org/abs/2605.29152v1
• Date: Wed, 27 May 2026 22:30:39 GMT
• Title: Do Deep Networks Forget Initialization? A Forgetting-Time View of Practical Inductive Bias
• Abstract summary: We show that low-learning-rate SGD can interpolate while still remembering its initialization.<n>We interpret these findings in terms of the time scale of forgetting: gradient-flow-like dynamics can preserve memory, whereas finite-step effects, explicit norm decay, and adaptive preconditioning erase it on scales governed by the size of explicit or implicit regularization.
• Score: 0.5405981353784005
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Randomly initialized neural networks induce a prior over functions, but the predictor used in practice is produced only after training. We ask how much of this initial bias survives the training pipeline. To make the question measurable, we introduce initialization memory: the dependence of the validation-selected predictor on the scale of the random initialization. We perform controlled CIFAR-10 experiments on ResNets where initialization memory already sharply separates training regimes. Low-learning-rate SGD can interpolate while still remembering its initialization: on ResNet-9 with batch size $b=128$, test accuracy varies by $26.5$ percentage points across initialization scales despite $\ge99.5\%$ training accuracy. This is not undertraining: extending the same low-learning-rate regime to $5{,}000$ epochs leaves the spread essentially unchanged. In contrast, Adam-family methods largely erase the dependence. SGD can also be made to forget when larger learning rates are paired with explicit $L_2$ norm control. We interpret these findings in terms of the time scale of forgetting: gradient-flow-like dynamics can preserve initialization memory, whereas stochastic finite-step effects, explicit norm decay, and adaptive preconditioning erase it on scales governed by the size of explicit or implicit regularization. The practical inductive bias of a trained network is therefore not the architectural prior alone, but the architectural prior after being filtered by the forgetting dynamics of the training pipeline; and the same regularizers that improve generalization are precisely those that erase memory of initialization.

Related papers

• A Random-Matrix Criterion for Initializing Gated Recurrent Neural Networks [0.0]
  In "reservoir computing", the weights of a readout are learned linearly while the reservoir weights are fixed and largely determine the richness, stability and memory of the resulting dynamics.<n>In the infinite-width limit it has been shown that meaningful initializations are those sitting at an effective critical point of the randomly layer model.<n>Here we derive a simple criterion to estimate the critical $g_c$ for a broad class of recurrent architectures and we show that it closely tracks the gain at which a gated-RNN reservoir achieves peak performance on a chaotic forecasting task.
  arXiv  Detail & Related papers  (2026-05-11T14:35:22Z)
• A Theory of How Pretraining Shapes Inductive Bias in Fine-Tuning [51.505728136705564]
  We develop an analytical theory of the pretraining-fine-tuning pipeline in diagonal linear networks.<n>We find that different initialization choices place the network into four distinct fine-tuning regimes.<n>A smaller scale in earlier layers enables the network to both reuse and refine its features, leading to superior generalization.
  arXiv  Detail & Related papers  (2026-02-23T17:19:33Z)
• Learning effective pruning at initialization from iterative pruning [15.842658282636876]
  We present an end-to-end neural network-based PaI method to reduce training costs. Our approach outperforms existing methods in high-sparsity settings. As the first neural network-based PaI method, we conduct extensive experiments to validate the factors influencing this approach.
  arXiv  Detail & Related papers  (2024-08-27T03:17:52Z)
• Winner-Take-All Column Row Sampling for Memory Efficient Adaptation of Language Model [89.8764435351222]
  We propose a new family of unbiased estimators called WTA-CRS, for matrix production with reduced variance. Our work provides both theoretical and experimental evidence that, in the context of tuning transformers, our proposed estimators exhibit lower variance compared to existing ones.
  arXiv  Detail & Related papers  (2023-05-24T15:52:08Z)
• What to Prune and What Not to Prune at Initialization [0.0]
  Post-training dropout based approaches achieve high sparsity. Initialization pruning is more efficacious when it comes to scaling computation cost of the network. The goal is to achieve higher sparsity while preserving performance.
  arXiv  Detail & Related papers  (2022-09-06T03:48:10Z)
• Robust Training of Neural Networks using Scale Invariant Architectures [70.67803417918854]
  In contrast to SGD, adaptive gradient methods like Adam allow robust training of modern deep networks. We show that this general approach is robust to rescaling of parameter and loss. We design a scale invariant version of BERT, called SIBERT, which when trained simply by vanilla SGD achieves performance comparable to BERT trained by adaptive methods like Adam.
  arXiv  Detail & Related papers  (2022-02-02T11:58:56Z)
• GradInit: Learning to Initialize Neural Networks for Stable and Efficient Training [59.160154997555956]
  We present GradInit, an automated and architecture method for initializing neural networks. It is based on a simple agnostic; the variance of each network layer is adjusted so that a single step of SGD or Adam results in the smallest possible loss value. It also enables training the original Post-LN Transformer for machine translation without learning rate warmup.
  arXiv  Detail & Related papers  (2021-02-16T11:45:35Z)
• Direction Matters: On the Implicit Bias of Stochastic Gradient Descent with Moderate Learning Rate [105.62979485062756]
  This paper attempts to characterize the particular regularization effect of SGD in the moderate learning rate regime. We show that SGD converges along the large eigenvalue directions of the data matrix, while GD goes after the small eigenvalue directions.
  arXiv  Detail & Related papers  (2020-11-04T21:07:52Z)
• Implicit Bias in Deep Linear Classification: Initialization Scale vs Training Accuracy [71.25689267025244]
  We show how the transition is controlled by the relationship between the scale and how accurately we minimize the training loss. Our results indicate that some limit behaviors of gradient descent only kick in at ridiculous training accuracies.
  arXiv  Detail & Related papers  (2020-07-13T23:49:53Z)
• Fractional moment-preserving initialization schemes for training deep neural networks [1.14219428942199]
  A traditional approach to deep neural networks (DNNs) is to sample the network weights randomly for preserving the variance of pre-activations. In this paper, we show that weights and therefore pre-activations can be modeled with a heavy-tailed distribution. We show through numerical experiments that our schemes can improve the training and test performance.
  arXiv  Detail & Related papers  (2020-05-25T01:10:01Z)

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.