Research map / Optimization / Implicit bias and generalization

Changing the Training Data Distribution to Reduce Simplicity Bias Improves In-distribution Generalization

NeurIPS 2024 · Optimization › Implicit bias and generalization · approach: Stability-based generalization analysis · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proves SAM reduces simplicity bias in a two-layer CNN and proposes USEFUL, which clusters examples early in training and upsamples the slow-learnable cluster once, improving in-distribution generalization for SGD and SAM on image classification.

Problem

Whether modifying the training data distribution (rather than the optimizer alone) can improve in-distribution generalization by reducing the simplicity bias of gradient-based training.

Approach

Theory shows SAM learns fast- and slow-learnable features more uniformly than GD; USEFUL mimics this by clustering examples by early-training outputs and upsampling the slow-learnable cluster once (k=2) before restarting training with SGD or SAM.

Key results

• USEFUL with SAM and TrivialAugment reaches state-of-the-art test errors reported for ResNet18 on CIFAR10 (3.49% vs 4.04% for SAM+TA... as reported in Figure 4) and ResNet34 on CIFAR100 (17.72%).
• USEFUL reduces SGD test error on CIFAR10 from 5.07% to 4.23% at 1.3x training cost, compared with 4.04% for SAM at 2x cost.
• USEFUL consistently beats random upsampling of the same data for ResNet18 on CIFAR10 and CIFAR100.

Measured effects reported in the paper

Surprising findings

• On STL10, USEFUL boosts SGD enough to surpass SAM's performance.
• Applying USEFUL to ViT-S and MLP lowers SGD error below that of SAM alone.

Limitations

• Theory is limited to a two-layer nonlinear CNN on a synthetic patch data distribution with orthogonal features.
• Upsampling is one-shot with a fixed factor k=2 and separating epoch, rather than adaptive multi-stage sampling.
• Out-of-distribution benefits are empirical only and not the paper's focus.

Experimental setup

Datasets: CIFAR10 CIFAR100 STL10 CINIC10 Tiny-ImageNet Waterbirds

Benchmarks: CIFAR10 CIFAR100 STL10 CINIC10 Tiny-ImageNet Waterbirds

Metrics: test classification error L1 norm maximum Hessian eigenvalue bulk of Hessian spectrum forgetting score

Baselines: SGD SAM SAM+TA ASAM random upsampling

Builds on: sharpness-aware minimization TrivialAugment ASAM

Closest papers by meaning

• Friendly Sharpness-Aware Minimization (CVPR 2024 · similarity 0.91)
• Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training (ICLR 2025 · similarity 0.91)
• Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning (ICLR 2024 · similarity 0.90)
• Simplicity Bias in Overparameterized Machine Learning (AAAI 2024 · similarity 0.89)
• Simplicity Bias and Optimization Threshold in Two-Layer ReLU Networks (ICML 2025 · similarity 0.89)
• A Single-Step, Sharpness-Aware Minimization is All You Need to Achieve Efficient and Accurate Sparse Training (NeurIPS 2024 · similarity 0.89)