What Do Compressed Deep Neural Networks Forget?

Abstract

Deep neural network pruning and quantization techniques have demonstrated it is possible to achieve high levels of compression with surprisingly little degradation to test set accuracy. However, this measure of performance conceals significant differences in how different classes and images are impacted by model compression techniques. We find that models with radically different numbers of weights have comparable top-line performance metrics but diverge considerably in behavior on a narrow subset of the dataset. This small subset of data points, which we term Pruning Identified Exemplars (PIEs) are systematically more impacted by the introduction of sparsity. Compression disproportionately impacts model performance on the underrepresented long-tail of the data distribution. PIEs over-index on atypical or noisy images that are far more challenging for both humans and algorithms to classify. Our work provides intuition into the role of capacity in deep neural networks and the trade-offs incurred by compression. An understanding of this disparate impact is critical given the widespread deployment of compressed models in the wild.

References in corpus (16)

• Distilling the Knowledge in a Neural Network
• On Calibration of Modern Neural Networks
• Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations
• Deep Learning with Limited Numerical Precision
• To prune, or not to prune: exploring the efficacy of pruning for model compression
• Learning Structured Sparsity in Deep Neural Networks
• The State of Sparsity in Deep Neural Networks
• Learning Efficient Convolutional Networks through Network Slimming
• Memory Bounded Deep Convolutional Networks
• MLIR: A Compiler Infrastructure for the End of Moore's Law
• Rigging the Lottery: Making All Tickets Winners
• What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation
• Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images
• What is the State of Neural Network Pruning?
• Towards Compact and Robust Deep Neural Networks
• Keep the Gradients Flowing: Using Gradient Flow to Study Sparse Network Optimization

Cited by in corpus (31)

• Knowledge Distillation in Deep Learning and its Applications
• Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks
• Are we done with ImageNet?
• Characterising Bias in Compressed Models
• HYDRA: Pruning Adversarially Robust Neural Networks
• Lost in Pruning: The Effects of Pruning Neural Networks beyond Test Accuracy
• Randomness In Neural Network Training: Characterizing The Impact of Tooling
• Can Subnetwork Structure be the Key to Out-of-Distribution Generalization?
• Self-Damaging Contrastive Learning
• Robustness-Reinforced Knowledge Distillation with Correlation Distance and Network Pruning
• Pruning a restricted Boltzmann machine for quantum state reconstruction
• An investigation of structures responsible for gender bias in BERT and DistilBERT
• Who's responsible? Jointly quantifying the contribution of the learning algorithm and training data
• The Rich Get Richer: Disparate Impact of Semi-Supervised Learning
• ReSmooth: Detecting and Utilizing OOD Samples when Training with Data Augmentation
• Robustness in Compressed Neural Networks for Object Detection
• Towards Understanding Iterative Magnitude Pruning: Why Lottery Tickets Win
• Robustness Challenges in Model Distillation and Pruning for Natural Language Understanding
• PaCKD: Pattern-Clustered Knowledge Distillation for Compressing Memory Access Prediction Models
• A Winning Hand: Compressing Deep Networks Can Improve Out-Of-Distribution Robustness
• Simon Says: Evaluating and Mitigating Bias in Pruned Neural Networks with Knowledge Distillation
• Understanding the effect of sparsity on neural networks robustness
• When does loss-based prioritization fail?
• Troubleshooting Blind Image Quality Models in the Wild
• A Tale Of Two Long Tails
• Measure Twice, Cut Once: Quantifying Bias and Fairness in Deep Neural Networks
• BERTnesia: Investigating the capture and forgetting of knowledge in BERT
• Are Compressed Language Models Less Subgroup Robust?
• Robust error bounds for quantised and pruned neural networks
• Identifying and Exploiting Structures for Reliable Deep Learning
• An Underexplored Dilemma between Confidence and Calibration in Quantized Neural Networks