Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ChronoSelect: Robust Learning with Noisy Labels via Dynamics Temporal Memory

Jianchao Wang, Qingfeng Li, Pengcheng Zheng, Xiaorong Pu, Yazhou Ren

TL;DR

This work tackles learning with noisy labels by introducing ChronoSelect, a temporal-memory framework that compresses entire prediction histories into four dynamic memory units per sample via a sliding update. The method leverages dual-branch consistency and two temporal signatures to partition data into clean, boundary, and noisy subsets, enabling tailored supervision for each group. The authors prove convergence and stability of the memory under noise and demonstrate state-of-the-art performance on synthetic (CIFAR) and real-world (WebVision) benchmarks, highlighting robustness and threshold-free operation. Overall, ChronoSelect provides a principled, scalable approach to robust learning with noisy labels by exploiting temporal learning dynamics and memory-based trajectory analysis.

Abstract

Training deep neural networks on real-world datasets is often hampered by the presence of noisy labels, which can be memorized by over-parameterized models, leading to significant degradation in generalization performance. While existing methods for learning with noisy labels (LNL) have made considerable progress, they fundamentally suffer from static snapshot evaluations and fail to leverage the rich temporal dynamics of learning evolution. In this paper, we propose ChronoSelect (chrono denoting its temporal nature), a novel framework featuring an innovative four-stage memory architecture that compresses prediction history into compact temporal distributions. Our unique sliding update mechanism with controlled decay maintains only four dynamic memory units per sample, progressively emphasizing recent patterns while retaining essential historical knowledge. This enables precise three-way sample partitioning into clean, boundary, and noisy subsets through temporal trajectory analysis and dual-branch consistency. Theoretical guarantees prove the mechanism's convergence and stability under noisy conditions. Extensive experiments demonstrate ChronoSelect's state-of-the-art performance across synthetic and real-world benchmarks.

ChronoSelect: Robust Learning with Noisy Labels via Dynamics Temporal Memory

TL;DR

This work tackles learning with noisy labels by introducing ChronoSelect, a temporal-memory framework that compresses entire prediction histories into four dynamic memory units per sample via a sliding update. The method leverages dual-branch consistency and two temporal signatures to partition data into clean, boundary, and noisy subsets, enabling tailored supervision for each group. The authors prove convergence and stability of the memory under noise and demonstrate state-of-the-art performance on synthetic (CIFAR) and real-world (WebVision) benchmarks, highlighting robustness and threshold-free operation. Overall, ChronoSelect provides a principled, scalable approach to robust learning with noisy labels by exploiting temporal learning dynamics and memory-based trajectory analysis.

Abstract

Training deep neural networks on real-world datasets is often hampered by the presence of noisy labels, which can be memorized by over-parameterized models, leading to significant degradation in generalization performance. While existing methods for learning with noisy labels (LNL) have made considerable progress, they fundamentally suffer from static snapshot evaluations and fail to leverage the rich temporal dynamics of learning evolution. In this paper, we propose ChronoSelect (chrono denoting its temporal nature), a novel framework featuring an innovative four-stage memory architecture that compresses prediction history into compact temporal distributions. Our unique sliding update mechanism with controlled decay maintains only four dynamic memory units per sample, progressively emphasizing recent patterns while retaining essential historical knowledge. This enables precise three-way sample partitioning into clean, boundary, and noisy subsets through temporal trajectory analysis and dual-branch consistency. Theoretical guarantees prove the mechanism's convergence and stability under noisy conditions. Extensive experiments demonstrate ChronoSelect's state-of-the-art performance across synthetic and real-world benchmarks.

Paper Structure

This paper contains 15 sections, 2 theorems, 16 equations, 1 figure, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Memorization of DNNs
  4. Handling Noisy Labels via Sample Selection
  5. Proposed Method
  6. Preliminaries
  7. Temporal Memory Space Construction and Sliding Update
  8. Convergence and Stability Analysis
  9. Sample Selection Strategy with the Temporal Memory Space
  10. Semi-Supervised Learning via Sample Selection for Noisy Labels
  11. Experimental
  12. Controllable Noise Benchmarks
  13. Real-world Noise Benchmarks
  14. Ablation Study
  15. Conclusion

Key Result

theorem 1

As training progresses ($t \to \infty$), each memory unit converges to the steady-state prediction $p^*$ when model predictions stabilize:

Figures (1)

  • Figure 1: (a) The framework of ChronoSelect. At each epoch, the ChronoSelect first employs a two-view-dual network for image classification, in which two views are obtained via two diverse augmentations. Then a dynamic sliding update strategy is designed to memorize the distribution of previous all epochs, which is used to divide the noisy data into tree subsets, i.e., clean set $\mathcal{D}_c$, boundary set $\mathcal{D}_b$ and noisy set $\mathcal{D}_n$ with consistency metric by dual network. Finally, we adopt different loss function to update the dual network. (b) The structure of Temporal Memory Space (TMS), which stores the historical prediction of each sample via sliding update.

Theorems & Definitions (2)

  • theorem 1: Memory Convergence
  • theorem 2: Memory Stability