Compression-aware Training of Deep Networks

Abstract

In recent years, great progress has been made in a variety of application domains thanks to the development of increasingly deeper neural networks. Unfortunately, the huge number of units of these networks makes them expensive both computationally and memory-wise. To overcome this, exploiting the fact that deep networks are over-parametrized, several compression strategies have been proposed. These methods, however, typically start from a network that has been trained in a standard manner, without considering such a future compression. In this paper, we propose to explicitly account for compression in the training process. To this end, we introduce a regularizer that encourages the parameter matrix of each layer to have low rank during training. We show that accounting for compression during training allows us to learn much more compact, yet at least as effective, models than state-of-the-art compression techniques.

Accepted at NIPS 2017

References in corpus (15)

• Adam: A Method for Stochastic Optimization
• Distilling the Knowledge in a Neural Network
• ADADELTA: An Adaptive Learning Rate Method
• Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)
• Exploiting Linear Structure Within Convolutional Networks for Efficient Evaluation
• Weight Normalization: A Simple Reparameterization to Accelerate Training of Deep Neural Networks
• Speeding up Convolutional Neural Networks with Low Rank Expansions
• Layer Normalization
• Learning Structured Sparsity in Deep Neural Networks
• Convolutional neural networks with low-rank regularization
• Memory Bounded Deep Convolutional Networks
• DecomposeMe: Simplifying ConvNets for End-to-End Learning
• NoiseOut: A Simple Way to Prune Neural Networks
• Learning Convolutional Neural Networks using Hybrid Orthogonal Projection and Estimation
• Hybrid Orthogonal Projection and Estimation (HOPE): A New Framework to Probe and Learn Neural Networks

Cited by in corpus (34)

• Exploring the Connection Between Binary and Spiking Neural Networks
• A Low Effort Approach to Structured CNN Design Using PCA
• ASAP: Architecture Search, Anneal and Prune
• Constructing Energy-efficient Mixed-precision Neural Networks through Principal Component Analysis for Edge Intelligence
• ExpandNets: Linear Over-parameterization to Train Compact Convolutional Networks
• AutoPruner: An End-to-End Trainable Filter Pruning Method for Efficient Deep Model Inference
• Mixed Dimension Embeddings with Application to Memory-Efficient Recommendation Systems
• Provable Filter Pruning for Efficient Neural Networks
• Data-Independent Neural Pruning via Coresets
• FRILL: A Non-Semantic Speech Embedding for Mobile Devices
• Group Sparsity: The Hinge Between Filter Pruning and Decomposition for Network Compression
• Mixed-Precision Embedding Using a Cache
• SiPPing Neural Networks: Sensitivity-informed Provable Pruning of Neural Networks
• Bringing Giant Neural Networks Down to Earth with Unlabeled Data
• Riemannian Low-Rank Model Compression for Federated Learning with Over-the-Air Aggregation
• Robust Pruning at Initialization
• DHP: Differentiable Meta Pruning via HyperNetworks
• Full deep neural network training on a pruned weight budget
• Sparse Weight Activation Training
• Smallify: Learning Network Size while Training
• On Compressing U-net Using Knowledge Distillation
• Data-Dependent Coresets for Compressing Neural Networks with Applications to Generalization Bounds
• Dynamic Sparse Graph for Efficient Deep Learning
• Compressing Neural Networks: Towards Determining the Optimal Layer-wise Decomposition
• VACL: Variance-Aware Cross-Layer Regularization for Pruning Deep Residual Networks
• Refining the Structure of Neural Networks Using Matrix Conditioning
• A Unified Approximation Framework for Compressing and Accelerating Deep Neural Networks
• Training Sparse Neural Networks using Compressed Sensing
• Pruning Ternary Quantization
• Low-memory convolutional neural networks through incremental depth-first processing
• Softer Pruning, Incremental Regularization
• Distilling Image Classifiers in Object Detectors
• HALO: Learning to Prune Neural Networks with Shrinkage
• A Highly Effective Low-Rank Compression of Deep Neural Networks with Modified Beam-Search and Modified Stable Rank