Self-Adaptive Training: beyond Empirical Risk Minimization

Lang Huang, Chao Zhang, Hongyang Zhang

TL;DR

Self-Adaptive Training is a simple yet powerful training algorithm that prevents deep neural networks from memorizing "garbage" data. By using a model’s own evolving predictions to dynamically correct labels and re-weight samples, it significantly improves performance on noisy or adversarial datasets without adding extra computational cost.

What Problem Does This Paper Address?

In the ideal world of machine learning, our training labels are perfect. In the real world, they are often "noisy"—corrupted by human error, sensor malfunctions, or even intentional adversarial attacks.

Standard Empirical Risk Minimization (ERM)—the bread and butter of deep learning—has a major flaw: deep networks are "universal approximators" that are too good at their jobs. If you give a powerful network a dataset where 40% of the labels are completely random, ERM will eventually force the network to memorize those random labels. It will achieve 100% training accuracy but fail miserably on actual test data.

Accuracy curves comparing standard ERM (top) which overfits to noise, and Self-Adaptive Training (bottom) which resists memorization.

This paper seeks a way to "filter" this noise during the training process itself, allowing the model to focus on the underlying signal rather than the corruptions.

Methods

The authors propose Self-Adaptive Training, which modifies the standard training loop with two clever mechanisms:

  1. Target Calibration (Moving Average): Instead of treating the initial (potentially noisy) label as the absolute truth, the algorithm maintains a "soft" target for each image. For the first few epochs (the "warm-up" phase), it uses the original labels. After that, it updates the target using an Exponential Moving Average (EMA) of the model’s own predictions. If the model consistently thinks an image labeled "Cat" is actually a "Dog," the target label eventually shifts toward "Dog."
  2. Sample Re-weighting: The algorithm calculates a weight for every sample based on the model's confidence (the maximum value in the soft target vector). If the model is highly uncertain about a sample, that sample’s influence on the total loss is diminished. This prevents the model from being "pulled" too hard by ambiguous or heavily corrupted data.

Crucially, this approach requires no changes to the network architecture and adds virtually zero computational overhead.

Datasets

The researchers evaluated their method across several standard benchmarks, often injecting synthetic noise to test robustness:

  • CIFAR-10 & CIFAR-100: Used with four types of noise: corrupted labels, Gaussian noise, random pixels, and shuffled pixels.
  • ImageNet: Tested on both the standard version (which contains inherent natural noise) and a version with 40% synthetic label noise.
  • Dogs vs. Cats: Used specifically for the "selective classification" task.

Key Insights and Ideas

  • The "Blessing" of Early Predictions: The authors observed that even when training on noisy data, a model learns useful, clean features in the first few epochs before it starts memorizing the noise. Self-adaptive training "captures" this early distilled knowledge to correct the labels before the overfitting phase begins.
  • Debunking "Double Descent": Recent AI research has been fascinated by "double descent," where test error goes down, then up, then down again as model capacity increases. The authors show that under self-adaptive training, this phenomenon disappears, resulting in a traditional "single descent" curve. This suggests that double descent might simply be a byproduct of ERM overfitting to noise.
  • Adversarial Robustness: The method isn't just for random noise. When applied to adversarial training (like the TRADES algorithm), it prevents "robust overfitting," where a model becomes less resistant to attacks the longer you train it.

Key Results

Self-Adaptive Training demonstrates remarkable resilience across various noise types and intensities, consistently outperforming standard ERM.

Generalization error across different randomization probabilities. Self-Adaptive Training (solid lines) maintains low error compared to ERM (dashed lines).
  • Label Noise: On CIFAR-100 with 80% noise, self-adaptive training outperformed the previous state-of-the-art by a massive 9.3% absolute accuracy margin.
  • ImageNet: Even on a massive scale, the method improved Top-1 accuracy by 2% under 40% label noise compared to standard ERM.
  • Selective Classification: In "selective classification" (where a model can choose to say "I don't know" rather than guess), the method achieved up to a 50% relative improvement in error rates.
  • Label Recovery: The algorithm is remarkably good at "fixing" the data. On CIFAR-10 with 40% noise, it successfully recovered the true labels with 94.6% accuracy.

Limitations and Future Work

  • Model Capacity: The authors note that the method can perform worse than ERM if the model is extremely small (underfitting). If the model isn't powerful enough to learn the "signal" during the warm-up phase, its predictions won't be reliable enough to correct the labels.
  • Hyperparameter Sensitivity: While the authors found the method robust, it still introduces two new hyperparameters: the warm-up duration ($E_s$) and the momentum term ($\alpha$). Finding the optimal "start time" for adaptation is critical.
  • Complex Noise: While the paper covers random and adversarial noise, future work could explore how this fares against "systematic" human bias in labeling, where errors aren't random but follow specific, misleading patterns.

Summary updated

Abstract

We propose self-adaptive training---a new training algorithm that dynamically calibrates training process by model predictions without incurring extra computational cost---to improve generalization of deep learning for potentially corrupted training data. This problem is important to robustly learning from data that are corrupted by, e.g., random noises and adversarial examples. The standard empirical risk minimization (ERM) for such data, however, may easily overfit noises and thus suffers from sub-optimal performance. In this paper, we observe that model predictions can substantially benefit the training process: self-adaptive training significantly mitigates the overfitting issue and improves generalization over ERM under both random and adversarial noises. Besides, in sharp contrast to the recently-discovered double-descent phenomenon in ERM, self-adaptive training exhibits a single-descent error-capacity curve, indicating that such a phenomenon might be a result of overfitting of noises. Experiments on the CIFAR and ImageNet datasets verify the effectiveness of our approach in two applications: classification with label noise and selective classification.
BibTeX
@inproceedings{huang2020self,
  author = {Lang Huang and Chao Zhang and Hongyang Zhang},
  title = {{Self-Adaptive Training: beyond Empirical Risk Minimization}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=nWwhYNoJ4C},
}
Figure