A Topological Filter for Learning with Label Noise

Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris N. Metaxas, Chao Chen

Abstract

Noisy labels can impair the performance of deep neural networks. To tackle this problem, in this paper, we propose a new method for filtering label noise. Unlike most existing methods relying on the posterior probability of a noisy classifier, we focus on the much richer spatial behavior of data in the latent representational space. By leveraging the high-order topological information of data, we are able to collect most of the clean data and train a high-quality model. Theoretically we prove that this topological approach is guaranteed to collect the clean data with high probability. Empirical results show that our method outperforms the state-of-the-arts and is robust to a broad spectrum of noise types and levels.

Introduction

Deep neural networks (DNNs) are highly effective but notoriously susceptible to memorizing training data, including incorrect or "noisy" labels. This memorization significantly degrades a model's ability to generalize to new data. Traditional approaches to label noise filtering often rely on classifier confidence—assuming that samples the model is "sure" about are likely clean. However, these heuristics can fail when noise patterns are complex or when models become overconfident in their own errors.

This paper introduces TopoFilter, a novel approach that shifts the focus from classifier probability to the spatial distribution of data in the latent feature space. By analyzing the topological structure of the penultimate layer's representations, TopoFilter identifies clean samples based on their tendency to form dense, connected clusters.

The Topological Perspective

The core intuition behind TopoFilter is that in a robust feature representation, clean samples of the same class cluster together into a large, dense "core," while noisy samples appear as isolated "crumbs" or small, scattered clusters.

Comparison of ideal, skewed, and learned representations in the latent space.

As shown in the figure above, an ideal representation (a) exhibits clear class separation. Noise introduces skew (b), but the underlying manifold of clean data remains more connected than the scattered noise. Through iterative training, TopoFilter recovers a clean representation (c) by focusing on these global topological features rather than just local density or individual sample probabilities.

The TopoFilter Algorithm

TopoFilter operates through an iterative process consisting of two primary components:

  1. TopoCC (Connected Components): For each class, the algorithm constructs a $k$-nearest neighbor (k-NN) graph in the latent space. It identifies the largest connected component as the primary candidate set for clean data. This captures the global structure of the class, effectively ignoring isolated noisy outliers.
  2. $\zeta$-filtering (The "Peeling" Step): To further refine the selection, TopoFilter "peels" the outer layers of the identified component. It requires that at least a fraction ($\zeta$) of a sample's neighbors share the same label, ensuring that noisy samples attached to the edges of the clean core are removed.
Illustration of the topological lemmas guiding the filtering process.

The effectiveness of this approach is grounded in several topological lemmas, visualized above:

  • Connectivity (Lemma 2): Clean samples tend to stay connected in the k-NN graph.
  • Isolation (Lemma 3): Noisy samples are likely to be isolated from the main clean component.
  • Kicking Out (Lemma 4): The filtering process effectively "kicks out" noisy points that might otherwise contaminate the clean set.

Theoretical and Empirical Results

Unlike many "black-box" heuristics, TopoFilter is supported by mathematical proofs for Purity (the collected data has a high probability of being clean) and Abundancy (the algorithm retains a significant portion of the clean data).

Empirically, TopoFilter sets new benchmarks across various datasets:

  • Extreme Noise Robustness: On CIFAR-10 with 80% uniform noise, TopoFilter achieved 45.7% accuracy, significantly outperforming baselines that often collapse under such high noise levels.
  • Real-World Performance: On the Clothing1M dataset, which contains naturally occurring "messy" labels, TopoFilter reached 74.10% accuracy, establishing a new state-of-the-art at the time of publication.
  • Stability: Unlike methods that suffer from performance drops as the model begins to memorize noise, TopoFilter maintains stable validation accuracy throughout the training process.

Conclusion

TopoFilter demonstrates that the spatial behavior of data in latent space is a powerful signal for noise filtering. By leveraging high-order topological information, it provides a robust and theoretically grounded alternative to probability-based methods. Future work may explore more advanced topological tools, such as Persistent Homology, to create differentiable losses that further automate the formation of clean data clusters.

Summary updated

Abstract

Noisy labels can impair the performance of deep neural networks. To tackle this problem, in this paper, we propose a new method for filtering label noise. Unlike most existing methods relying on the posterior probability of a noisy classifier, we focus on the much richer spatial behavior of data in the latent representational space. By leveraging the high-order topological information of data, we are able to collect most of the clean data and train a high-quality model. Theoretically we prove that this topological approach is guaranteed to collect the clean data with high probability. Empirical results show that our method outperforms the state-of-the-arts and is robust to a broad spectrum of noise types and levels.
BibTeX
@inproceedings{wu2020topological,
  author = {Pengxiang Wu and Songzhu Zheng and Mayank Goswami and Dimitris N. Metaxas and Chao Chen},
  title = {{A Topological Filter for Learning with Label Noise}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=rmroNgOhnJ5},
}
Figure