This is a collection of papers aiming at reducing model sizes or the ASIC/FPGA accelerator for Machine Learning, especially deep neural network related applications. (Inspired by Embedded-Neural-Network.)

You can use the following materials as your entrypoint:

• Efficient Processing of Deep Neural Networks: A Tutorial and Survey
• the related work of Quantized Neural Networks

Terminologies

• Structural pruning (compression): compress CNNs based on removing "less important" filter.

Network Compression

Reduce Precision

Deep neural networks are robust to weight binarization and other non-linear distortions showed that DNN can be robust to more than just weight binarization.

Linear Quantization

• Fixed point
  • [1502]. Deep Learning with Limited Numerical Precision
  • [1610]. QSGD: Communication-Optimal Stochastic Gradient Descent, with Applications to Training Neural Networks
• Dynamic fixed point
  • [1412]. Training deep neural networks with low precision multiplications
  • [1604]. Hardware-oriented approximation of convolutional neural networks
  • [1608]. Scalable and modularized RTL compilation of convolutional neural networks onto FPGA
• Binary Quantization
  • Theory proof (EBP)
    • [1405]. Expectation Backpropagation: Parameter-Free Training of Multilayer Neural Networks with Continuous or Discrete Weights
    • [1503]. Training Binary Multilayer Neural Networks for Image Classification using Expectation Backpropagation
    • [1505]. Backpropagation for Energy-Efficient Neuromorphic Computing
  • More practice with 1 bit
    • [1511]. BinaryConnect: Training Deep Neural Networks with binary weights during propagations
    • [1510]. Neural Networks with Few Multiplications
    • [1601]. Bitwise Neural Networks
    • [1602]. Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1
    • [1603]. XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks
  • XNOR-Net with slightly large bits (1~2 bit)
    • [1606]. DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients
    • [1608]. Recurrent Neural Networks With Limited Numerical Precision
    • [1609]. Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations. (Text overlap with Binarized Neural Network.)
    • [1702]. Deep Learning with Low Precision by Half-wave Gaussian Quantization
• Ternary Quantization
  • [1410]. Fixed-point feedforward deep neural network design using weights +1, 0, and -1
  • [1605]. Ternary Weight Networks
  • [1612]. Trained Ternary Quantization
• Other Quantization or others
  • [1412]. Compressing Deep Convolutional Networks using Vector Quantization
• 1-Bit Stochastic Gradient Descent and its Application to Data-Parallel Distributed Training of Speech DNNs
• Towards the Limit of Network Quantization.
• Loss-aware Binarization of Deep Networks.

Non-linear Quantization

• Log Domain Quantization
  • [1603]. Convolutional neural networks using logarithmic data representation
  • [1609]. LogNet: Energy-efficient neural networks using logarithmic computation
  • [1702]. Incremental Network Quantization: Towards Lossless CNNs with Low-Precision Weights
• Parameter Sharing
  • Structured Matrices
    • Structured Convolution Matrices for Energy-efficient Deep learning.
    • Structured Transforms for Small-Footprint Deep Learning.
    • An Exploration of Parameter Redundancy in Deep Networks with Circulant Projections.
    • Theoretical Properties for Neural Networks with Weight Matrices of Low Displacement Rank.
  • Hashing
    • [1504]. Compressing neural networks with the hashing trick
    • Functional Hashing for Compressing Neural Networks
  • [1510]. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding
  • Learning compact recurrent neural networks.

Reduce Number of Operations and Model Size

Exploiting Activation Statistics

• To be updated.

Network Pruning

Network Prune: a large amount of the weights in a network are redundant and can be removed (i.e., set to zero).

• Remove low saliency
  • [9006]. Optimal Brain Damage
  • [1506]. Learning both weights and connections for efficient neural network
• Energy-based prune
  • [1611]. Designing energy-efficient convolutional neural networks using energy-aware pruning
• Process sparse weights
  • [1402]. A scalable sparse matrix-vector multiplication kernel for energy-efficient sparse-blas on FPGAs
  • [1510]. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding
  • [1602]. EIE: Efficient Inference Engine on Compressed Deep Neural Network
  • [1705]. SCNN: An Accelerator for Compressed-sparse Convolutional Neural Networks
  • [1710]. Efficient Methods and Hardware for Deep Learning, Ph.D. Thesis
• Structured pruning
  • [1512]. Structured Pruning of Deep Convolutional Neural Networks
  • [1608]. Learning Structured Sparsity in Deep Neural Networks
  • [1705]. Exploring the Regularity of Sparse Structure in Convolutional Neural Networks

Bayesian network pruning

• [1711]. Interpreting Convolutional Neural Networks Through Compression - [notes][arXiv]
• [1705]. Structural compression of convolutional neural networks based on greedy filter pruning - [notes][arXiv]

Compact Network Architectures

• Before Training
  • use 1*1 convolutional layer to reduce the number of channels
    • [1512]. Rethinking the inception architecture for computer vision
    • [1610]. Xception: Deep Learning with Depthwise Separable Convolutions
    • [1704]. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
  • Bottleneck:
    • [1312]. Network in network
    • [1409]. Going deeper with convolutions
    • [1512]. Deep residual learning for image recognition
    • [1602]. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size
• After Training
  • Canonical Polyadic (CP) decomposition
    • [1404]. Exploiting linear structure within convolutional networks for efficient evaluation
    • [1412]. Speeding-up convolutional neural networks using fine-tuned cp-decomposition
  • Tucker decomposition
    • [1511]. Compression of deep convolutional neural networks for fast and low power mobile applications

Knowledge Distillation

• [0600]. Model compression
• [1312]. Do deep nets really need to be deep?
• [1412]. Fitnets: Hints for thin deep nets
• [1503]. Distilling the knowledge in a neural network
• Sequence-Level Knowledge Distillation.
• Like What You Like: Knowledge Distill via Neuron Selectivity Transfer.

A Bit Hardware

• [1402]. Computing's Energy Porblem (and what we can do about it)

Contributors

• Tao Lin
• Jun Lu