Implicit Bias of SignGD and Adam on Multiclass Separable Data

• URL: http://arxiv.org/abs/2502.04664v1
• Date: Fri, 07 Feb 2025 05:09:32 GMT
• Title: Implicit Bias of SignGD and Adam on Multiclass Separable Data
• Authors: Chen Fan, Mark Schmidt, Christos Thrampoulidis,
• Abstract summary: Different gradient-based methods can achieve zero training error yet converge to distinctly different solutions inducing different generalization properties. We characterize the implicit bias of both Adam and Sign Gradient Descent in multi-class cross-entropy minimization. We extend our results to general p-norm normalized steepest descent algorithms and to other multi-class losses.
• Score: 33.082961718280245
• License:
• Abstract: In the optimization of overparameterized models, different gradient-based methods can achieve zero training error yet converge to distinctly different solutions inducing different generalization properties. While a decade of research on implicit optimization bias has illuminated this phenomenon in various settings, even the foundational case of linear classification with separable data still has important open questions. We resolve a fundamental gap by characterizing the implicit bias of both Adam and Sign Gradient Descent in multi-class cross-entropy minimization: we prove that their iterates converge to solutions that maximize the margin with respect to the classifier matrix's max-norm and characterize the rate of convergence. We extend our results to general p-norm normalized steepest descent algorithms and to other multi-class losses.

Related papers

• Differentially Private Optimization with Sparse Gradients [60.853074897282625]
  We study differentially private (DP) optimization problems under sparsity of individual gradients. Building on this, we obtain pure- and approximate-DP algorithms with almost optimal rates for convex optimization with sparse gradients.
  arXiv  Detail & Related papers  (2024-04-16T20:01:10Z)
• Precise Asymptotic Generalization for Multiclass Classification with Overparameterized Linear Models [4.093769373833101]
  We resolve the conjecture posed in Subramanian et al.'22, where the number of data points, features, and classes all grow together. Our new lower bounds are akin to an information-theoretic strong converse: they establish that the misclassification rate goes to 0 or 1ally. The key to our tight analysis is a new variant of the Hanson-Wright inequality which is broadly useful for multiclass problems with sparse labels.
  arXiv  Detail & Related papers  (2023-06-23T00:59:15Z)
• Stability vs Implicit Bias of Gradient Methods on Separable Data and Beyond [33.593203156666746]
  We focus on the generalization properties of unregularized gradient-based learning procedures applied to separable linear classification. We give an additional unified explanation for this generalization, that we refer to as realizability and self-boundedness. In some of these cases, our bounds significantly improve upon the existing generalization error bounds in the literature.
  arXiv  Detail & Related papers  (2022-02-27T19:56:36Z)
• Improving Generalization via Uncertainty Driven Perturbations [107.45752065285821]
  We consider uncertainty-driven perturbations of the training data points. Unlike loss-driven perturbations, uncertainty-guided perturbations do not cross the decision boundary. We show that UDP is guaranteed to achieve the robustness margin decision on linear models.
  arXiv  Detail & Related papers  (2022-02-11T16:22:08Z)
• Sharp global convergence guarantees for iterative nonconvex optimization: A Gaussian process perspective [30.524043513721168]
  We develop a general recipe for analyzing the convergence of iterative algorithms for a class of regression models. deterministicly, we accurately capture both the convergence rate of the algorithm and the eventual error floor in the finite-sample regime. We show sharp convergence rates for both higher-order algorithms based on alternating updates and first-order algorithms based on subgradient subgradients.
  arXiv  Detail & Related papers  (2021-09-20T21:48:19Z)
• Understanding the Generalization of Adam in Learning Neural Networks with Proper Regularization [118.50301177912381]
  We show that Adam can converge to different solutions of the objective with provably different errors, even with weight decay globalization. We show that if convex, and the weight decay regularization is employed, any optimization algorithms including Adam will converge to the same solution.
  arXiv  Detail & Related papers  (2021-08-25T17:58:21Z)
• Learning Gaussian Mixtures with Generalised Linear Models: Precise Asymptotics in High-dimensions [79.35722941720734]
  Generalised linear models for multi-class classification problems are one of the fundamental building blocks of modern machine learning tasks. We prove exacts characterising the estimator in high-dimensions via empirical risk minimisation. We discuss how our theory can be applied beyond the scope of synthetic data.
  arXiv  Detail & Related papers  (2021-06-07T16:53:56Z)
• Benign Overfitting of Constant-Stepsize SGD for Linear Regression [122.70478935214128]
  inductive biases are central in preventing overfitting empirically. This work considers this issue in arguably the most basic setting: constant-stepsize SGD for linear regression. We reflect on a number of notable differences between the algorithmic regularization afforded by (unregularized) SGD in comparison to ordinary least squares.
  arXiv  Detail & Related papers  (2021-03-23T17:15:53Z)
• To Each Optimizer a Norm, To Each Norm its Generalization [31.682969645989512]
  We study the implicit regularization of optimization methods for linear models interpolating the training data in the under-parametrized and over-parametrized regimes. We argue that analyzing convergence to the standard maximum l2-margin is arbitrary and show that minimizing the norm induced by the data results in better generalizations.
  arXiv  Detail & Related papers  (2020-06-11T21:07:38Z)
• Asymptotic Analysis of an Ensemble of Randomly Projected Linear Discriminants [94.46276668068327]
  In [1], an ensemble of randomly projected linear discriminants is used to classify datasets. We develop a consistent estimator of the misclassification probability as an alternative to the computationally-costly cross-validation estimator. We also demonstrate the use of our estimator for tuning the projection dimension on both real and synthetic data.
  arXiv  Detail & Related papers  (2020-04-17T12:47:04Z)

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.