The Optimization Landscape of SGD Across the Feature Learning Strength

• URL: http://arxiv.org/abs/2410.04642v2
• Date: Tue, 8 Oct 2024 12:28:22 GMT
• Title: The Optimization Landscape of SGD Across the Feature Learning Strength
• Authors: Alexander Atanasov, Alexandru Meterez, James B. Simon, Cengiz Pehlevan,
• Abstract summary: We study the effect of scaling $gamma$ across a variety of models and datasets in the online training setting. We find that optimal online performance is often found at large $gamma$. Our findings indicate that analytical study of the large-$gamma$ limit may yield useful insights into the dynamics of representation learning in performant models.
• Score: 102.1353410293931
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: We consider neural networks (NNs) where the final layer is down-scaled by a fixed hyperparameter $\gamma$. Recent work has identified $\gamma$ as controlling the strength of feature learning. As $\gamma$ increases, network evolution changes from "lazy" kernel dynamics to "rich" feature-learning dynamics, with a host of associated benefits including improved performance on common tasks. In this work, we conduct a thorough empirical investigation of the effect of scaling $\gamma$ across a variety of models and datasets in the online training setting. We first examine the interaction of $\gamma$ with the learning rate $\eta$, identifying several scaling regimes in the $\gamma$-$\eta$ plane which we explain theoretically using a simple model. We find that the optimal learning rate $\eta^*$ scales non-trivially with $\gamma$. In particular, $\eta^* \propto \gamma^2$ when $\gamma \ll 1$ and $\eta^* \propto \gamma^{2/L}$ when $\gamma \gg 1$ for a feed-forward network of depth $L$. Using this optimal learning rate scaling, we proceed with an empirical study of the under-explored "ultra-rich" $\gamma \gg 1$ regime. We find that networks in this regime display characteristic loss curves, starting with a long plateau followed by a drop-off, sometimes followed by one or more additional staircase steps. We find networks of different large $\gamma$ values optimize along similar trajectories up to a reparameterization of time. We further find that optimal online performance is often found at large $\gamma$ and could be missed if this hyperparameter is not tuned. Our findings indicate that analytical study of the large-$\gamma$ limit may yield useful insights into the dynamics of representation learning in performant models.

Related papers

• Variance-reduced Clipping for Non-convex Optimization [24.765794811146144]
  Gradient clipping is a technique used in deep learning applications such as large-scale language modeling. Recent experimental training have a fairly special behavior in that it mitigates order complexity.
  arXiv  Detail & Related papers  (2023-03-02T00:57:38Z)
• The Onset of Variance-Limited Behavior for Networks in the Lazy and Rich Regimes [75.59720049837459]
  We study the transition from infinite-width behavior to this variance limited regime as a function of sample size $P$ and network width $N$. We find that finite-size effects can become relevant for very small datasets on the order of $P* sim sqrtN$ for regression with ReLU networks.
  arXiv  Detail & Related papers  (2022-12-23T04:48:04Z)
• Maximal Initial Learning Rates in Deep ReLU Networks [32.157430904535126]
  We introduce the maximal initial learning rate $etaast$. We observe that in constant-width fully-connected ReLU networks, $etaast$ behaves differently from the maximum learning rate later in training.
  arXiv  Detail & Related papers  (2022-12-14T15:58:37Z)
• Understanding Deep Neural Function Approximation in Reinforcement Learning via $\epsilon$-Greedy Exploration [53.90873926758026]
  This paper provides a theoretical study of deep neural function approximation in reinforcement learning (RL) We focus on the value based algorithm with the $epsilon$-greedy exploration via deep (and two-layer) neural networks endowed by Besov (and Barron) function spaces. Our analysis reformulates the temporal difference error in an $L2(mathrmdmu)$-integrable space over a certain averaged measure $mu$, and transforms it to a generalization problem under the non-iid setting.
  arXiv  Detail & Related papers  (2022-09-15T15:42:47Z)
• Neural Networks can Learn Representations with Gradient Descent [68.95262816363288]
  In specific regimes, neural networks trained by gradient descent behave like kernel methods. In practice, it is known that neural networks strongly outperform their associated kernels.
  arXiv  Detail & Related papers  (2022-06-30T09:24:02Z)
• High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation [89.21686761957383]
  We study the first gradient descent step on the first-layer parameters $boldsymbolW$ in a two-layer network. Our results demonstrate that even one step can lead to a considerable advantage over random features.
  arXiv  Detail & Related papers  (2022-05-03T12:09:59Z)
• Large-time asymptotics in deep learning [0.0]
  We consider the impact of the final time $T$ (which may indicate the depth of a corresponding ResNet) in training. For the classical $L2$--regularized empirical risk minimization problem, we show that the training error is at most of the order $mathcalOleft(frac1Tright)$. In the setting of $ellp$--distance losses, we prove that both the training error and the optimal parameters are at most of the order $mathcalOleft(e-mu
  arXiv  Detail & Related papers  (2020-08-06T07:33:17Z)
• On the Modularity of Hypernetworks [103.1147622394852]
  We show that for a structured target function, the overall number of trainable parameters in a hypernetwork is smaller by orders of magnitude than the number of trainable parameters of a standard neural network and an embedding method.
  arXiv  Detail & Related papers  (2020-02-23T22:51:52Z)

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.