Leveraging Dissimilarity Invariance as a Robust Anchor for Learning with Noisy Labels

2026-07-20

AI Summary

Deep learning performance degrades significantly under label noise, and existing methods are often misled due to the difficulty of accurately modeling sample similarity. This work is the first to identify that the “dissimilarity invariance” among semantically unrelated samples remains stable even in the presence of noise. Building on this insight, the authors propose NegScale, a plug-and-play framework that shifts the learning focus from noise-sensitive similarity to robust dissimilarity. NegScale integrates a Structured Negative Orthogonality Penalty (SNOP) and a Dissimilarity-Calibrated Similarity Adjustment (DCSA) mechanism, offering both theoretical grounding and practical training strategies. Extensive experiments demonstrate that NegScale substantially outperforms state-of-the-art methods on CIFAR benchmarks with synthetic noise as well as multiple real-world datasets, establishing a new performance standard.

Abstract

Deep learning models excel in visual recognition but suffer severe performance drops when training labels are corrupted by noise. Under label noise prior work cannot learn accurate similarities and thus misguide the learning process. In this paper, we uncover a complementary and novel phenomenon, Dissimilarity Invariance, whereby semantic dissimilarity between unrelated samples remains stable despite label noise. Leveraging this insight, we propose NegScale, a plug-and-play framework that shifts focus from fragile similarity to robust dissimilarity. NegScale integrates: (1) Structured Negative Orthogonality Penalty (SNOP), enforcing subspace orthogonality for unrelated samples; and (2) Dissimilarity-Calibrated Similarity Adjustment (DCSA), suppressing spurious similarity using dissimilarity anchors. We also give theoretical analysis that proves Dissimilarity Invariance and the effectiveness of NegScale. Empirical results demonstrate that NegScale consistently outperforms state-of-the-art baselines, establishing new benchmarks on CIFAR with synthetic noise and real-world datasets.

Problem

• noisy labels
• dissimilarity invariance
• semantic similarity
• robust learning
• label noise

Innovation

• Dissimilarity Invariance
• NegScale
• Noisy Labels
• Structured Negative Orthogonality Penalty
• Dissimilarity-Calibrated Similarity Adjustment