Towards Understanding Knowledge Distillation

Abstract

Knowledge distillation, i.e., one classifier being trained on the outputs of another classifier, is an empirically very successful technique for knowledge transfer between classifiers. It has even been observed that classifiers learn much faster and more reliably if trained with the outputs of another classifier as soft labels, instead of from ground truth data. So far, however, there is no satisfactory theoretical explanation of this phenomenon. In this work, we provide the first insights into the working mechanisms of distillation by studying the special case of linear and deep linear classifiers. Specifically, we prove a generalization bound that establishes fast convergence of the expected risk of a distillation-trained linear classifier. From the bound and its proof we extract three key factors that determine the success of distillation: * data geometry -- geometric properties of the data distribution, in particular class separation, has a direct influence on the convergence speed of the risk; * optimization bias -- gradient descent optimization finds a very favorable minimum of the distillation objective; and * strong monotonicity -- the expected risk of the student classifier always decreases when the size of the training set grows.

ICML'19. Post-edited to add related work. arXiv admin note: text overlap with arXiv:2003.13438 by other authors

Cited by in corpus (24)

• A Comprehensive Study of Class Incremental Learning Algorithms for Visual Tasks
• Rethinking Soft Labels for Knowledge Distillation: A Bias-Variance Tradeoff Perspective
• MalBERT: Using Transformers for Cybersecurity and Malicious Software Detection
• Why distillation helps: a statistical perspective
• Zero-Shot Knowledge Distillation from a Decision-Based Black-Box Model
• Transferring Inductive Biases through Knowledge Distillation
• An Empirical Analysis of the Impact of Data Augmentation on Knowledge Distillation
• Knowledge Distillation for Multi-task Learning
• Knowledge Distillation as Semiparametric Inference
• Dynamic Group Convolution for Accelerating Convolutional Neural Networks
• Compact representations of convolutional neural networks via weight pruning and quantization
• A Selective Survey on Versatile Knowledge Distillation Paradigm for Neural Network Models
• Robustness and Diversity Seeking Data-Free Knowledge Distillation
• On the Demystification of Knowledge Distillation: A Residual Network Perspective
• Deeplite Neutrino: An End-to-End Framework for Constrained Deep Learning Model Optimization
• Follow Your Path: a Progressive Method for Knowledge Distillation
• Embracing the Dark Knowledge: Domain Generalization Using Regularized Knowledge Distillation
• Temporal Self-Ensembling Teacher for Semi-Supervised Object Detection
• Transfer Heterogeneous Knowledge Among Peer-to-Peer Teammates: A Model Distillation Approach
• A Study of Non-autoregressive Model for Sequence Generation
• Understanding the Logit Distributions of Adversarially-Trained Deep Neural Networks
• Tensor train decompositions on recurrent networks
• Student-Teacher Learning from Clean Inputs to Noisy Inputs
• Lifelong Twin Generative Adversarial Networks