On Robustness of Neural Architecture Search Under Label Noise

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Neural Architecture Search (NAS) automates the design of neural networks for specific tasks.
• Real-world datasets often contain noisy labels due to human error or inadequate information.
• Label noise adversely affects neural network training and performance evaluation.

Purpose of the Study:

• To systematically investigate the robustness of Neural Architecture Search (NAS) under varying degrees of label noise.
• To evaluate the impact of label noise on both training and validation datasets in NAS.
• To explore methods for mitigating performance degradation caused by label noise in NAS.

Main Methods:

• Empirical experiments were conducted to assess NAS performance with noisy labels.
• Robust loss functions were employed to mitigate the effects of label noise.
• Theoretical analysis was performed to justify the observed empirical results.

Main Results:

• Label noise in training and/or validation data significantly impacts NAS performance.
• Robust loss functions demonstrated effectiveness in reducing performance degradation under symmetric and class-conditional label noise.
• Both empirical findings and theoretical justifications support the use of robust loss functions in noisy NAS scenarios.

Conclusions:

• Label noise poses a critical challenge for Neural Architecture Search (NAS).
• Employing robust loss functions is a viable strategy to enhance NAS resilience to label noise.
• The findings advocate for the integration of robust loss functions in NAS pipelines dealing with imperfect data.