Phase-Amplitude Spectrum Disentangled Early Stopping for Learning with Noisy Labels

Abstract

In recent years, convolutional neural networks (CNNs) have risen to prominence in vision tasks, demonstrating unmatched capabilities in pattern recognition and image classification. Despite their strengths, a persistent challenge is their vulnerability to label noise. When trained on datasets marred by mislabeling, CNNs often succumb to overfitting, which diminishes their performance on new, unseen data. A prevalent remedy for this issue is the early stopping strategy, which halts training before overfitting sets in, thereby preventing the model from assimilating the noise. The efficacy of early stopping can be further amplified when paired with insights from the biological vision system. Research in this domain has highlighted the unique roles of the amplitude spectrum (AS) and the phase spectrum (PS) in visual processing. Intriguingly, the phase spectrum, which encapsulates richer semantic information in images, has proven more potent in enhancing the resilience of CNNs to label noise than its amplitude counterpart. Inspired by these findings, we present the Phase-AmplituDe DisentangLed Early Stopping (PADDLES) method. This novel technique utilizes the discrete Fourier transform (DFT) to partition features into their respective amplitude and phase spectrum components. By judiciously applying early stopping at varied stages of training for each component, PADDLES capitalizes on the robust attributes of the phase spectrum while curbing the potential drawbacks of the amplitude spectrum. Through rigorous experimentation, PADDLES has showcased its effectiveness. Whether tested on synthetic datasets infused with artificial noise or real-world datasets with inherent mislabeling, PADDLES consistently surpasses conventional early stopping methods. Furthermore, it establishes new state-of-the-art benchmarks, redefining standards for training CNNs amidst label noise.