Do We Need Zero Training Loss After Achieving Zero Training Error?

Abstract

Overparameterized deep networks have the capacity to memorize training data with zero \emph{training error}. Even after memorization, the \emph{training loss} continues to approach zero, making the model overconfident and the test performance degraded. Since existing regularizers do not directly aim to avoid zero training loss, it is hard to tune their hyperparameters in order to maintain a fixed/preset level of training loss. We propose a direct solution called \emph{flooding} that intentionally prevents further reduction of the training loss when it reaches a reasonably small value, which we call the \emph{flood level}. Our approach makes the loss float around the flood level by doing mini-batched gradient descent as usual but gradient ascent if the training loss is below the flood level. This can be implemented with one line of code and is compatible with any stochastic optimizer and other regularizers. With flooding, the model will continue to "random walk" with the same non-zero training loss, and we expect it to drift into an area with a flat loss landscape that leads to better generalization. We experimentally show that flooding improves performance and, as a byproduct, induces a double descent curve of the test loss.

arXiv commentICML 2020 camera ready version

References in corpus (13)

• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• Decoupled Weight Decay Regularization
• Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
• Practical Bayesian Optimization of Machine Learning Algorithms
• On Calibration of Modern Neural Networks
• Reconciling modern machine learning practice and the bias-variance trade-off
• A Closer Look at Memorization in Deep Networks
• Dropout Training as Adaptive Regularization
• Augmenting Data with Mixup for Sentence Classification: An Empirical Study
• Overfitting or perfect fitting? Risk bounds for classification and regression rules that interpolate
• SIGUA: Forgetting May Make Learning with Noisy Labels More Robust
• SGD on Neural Networks Learns Functions of Increasing Complexity
• Mitigating Overfitting in Supervised Classification from Two Unlabeled Datasets: A Consistent Risk Correction Approach

Cited by in corpus (7)

• Kernel Based Progressive Distillation for Adder Neural Networks
• When Does Preconditioning Help or Hurt Generalization?
• One-Class Risk Estimation for One-Class Hyperspectral Image Classification
• Effective Regularization Through Loss-Function Metalearning
• Jitter: Random Jittering Loss Function
• Logit Attenuating Weight Normalization
• A Transformer Based Pitch Sequence Autoencoder with MIDI Augmentation