Towards Accurate Binary Convolutional Neural Network

Size: px
Start display at page:

Download "Towards Accurate Binary Convolutional Neural Network"

Transcription

1 Paper: #261 Poster: Pacific Ballroom #101 Towards Accurate Binary Convolutional Neural Network Xiaofan Lin, Cong Zhao and Wei Pan* Photos and videos are either original work or taken from Wikimedia, under Creative Commons license

2 DJI Drones use CNN for many tasks Large model size: Hundreds of megabytes of floating point weight values Expensive computation: Billions of floating point multiplication-accumulation Challenges for DJI Drones Limited resources for computation and power in DJI drones

3 Compression of deep neural networks for mobile applications Synapse and neuron pruning Quantization Sparse, irregular computation -- difficult to process efficiently Regular computation, smaller datapaths, fewer bits per weight and activation [1] S.Han, H.Mao, and J.W.Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arxiv preprint arxiv: , [2] P.Molchanov, S.Tyree, T.Karras, et al. Pruning Convolutional Neural Networks for Resource Efficient Inference. ICLR [3] I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. arxiv preprint arxiv: , [4] S. Zhou, Y. Wu, Z. Ni, X. Zhou, H. Wen, and Y. Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arxiv preprint arxiv: , [5] G.A.Howard, M.Zhu, B.Chen, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arxiv preprint arxiv: , [6] N.F.Iandola, S.Han, W.M.Moskewicz, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size. arxiv preprint arxiv: , 2016.

4 Compression of deep neural networks for mobile applications Synapse and neuron pruning Quantization Sparse, irregular computation -- difficult to process efficiently Regular computation, smaller datapaths, fewer bits per weight and activation [1] S.Han, H.Mao, and J.W.Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arxiv preprint arxiv: , [2] P.Molchanov, S.Tyree, T.Karras, et al. Pruning Convolutional Neural Networks for Resource Efficient Inference. ICLR [3] I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. arxiv preprint arxiv: , [4] S. Zhou, Y. Wu, Z. Ni, X. Zhou, H. Wen, and Y. Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arxiv preprint arxiv: , [5] G.A.Howard, M.Zhu, B.Chen, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arxiv preprint arxiv: , [6] N.F.Iandola, S.Han, W.M.Moskewicz, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size. arxiv preprint arxiv: , Floating point multiplication and accumulation is still the bottleneck!

5 Binarized neural networks (BNNs) The extreme case of quantization: binary weight and activation +1 and -1 (using sign() function) Key computation: binary matrix multiplication and accumulation x y = POPCOUNT(x XNOR y),x i,y i 2 { 1, +1}, 8i An example: 1 1 apple 1 1 Floating point operation 1 ( 1) + ( 1) 1 Bitwise operation! POPCOUNT(1 XNOR ( 1), 1 XNOR 1) XNOR Truth Table Input Output x y XNOR [7] M. Courbariaux, I. Hubara, D. Soudry, R. El-Yaniv, and Y. Bengio. Binarized neural networks: Training deep neural networks with weights and activations constrained to + 1 or-1, ICML [8] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. ECCV 2016.

6 Prediction accuracy with BNNs Competitive on small benchmarks: MNIST (handwritten digits), SVHN (street number), CIFAR-10 (10 classes objects) Too much loss on large benchmarks: ImageNet (1000 classes objects). MNIST SVHN CIFAR-10 ImageNet Binary weights & activations 99.04% 97.47% 89.85% 51.2% Full Precision weights & activations 99.06% 98.31% 92.38% 69.3% Accuracy loss 0.2% 0.84% 2.53% 18.1% [8] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. ECCV 2016.

7 Binary matrix multiplication Observation: too much accuracy loss using sign() for binarization Real-Value Weights Binary Weights Real-Value Inputs Binary Inputs Top-1 accuracy: 69.3% 60.8% Top-1 accuracy: 69.3% 51.2% Plan: approximate the weights and activations more precisely Intuitive example: say, we want to approximate a real number x =1.512 f 1 (x) =sign(x) ) x>0 f 1 (x) =sign(x),f 2 (x) = sign(x 1) ) x>1 f 1 (x) =sign(x),f 2 (x) = sign(x 1),f 3 (x) = sign(x 2) ) 1 <x<2 base 1 base 2 base 3

8 Approximate full precision weights using shift parameters Construct binary wight bases by shift B 1, B 2,, B M 2 { 1, +1} w h c in c out B i = F ui (W ) := sign(w mean(w )+u i std(w )) move the weight in sign() function by certain shift parameters shift parameters can be learned Approximate full precision weights using binary bases W 1 B B M B M

9 Approximate full precision activations using shift parameters Construct binary activation bases by shift 1 R = ReLU(Input) 2 h v (R) =clip(r + v, 0, 1) 3 H v (R) :=2I hv (R) Approximate full precision activations using binary bases A 1,, A N = H v1 (R),,H vn (R) R 1 A N A N

10 Approximate full precision activations by shift Construct binary activation bases by shift 1 R = ReLU(Input) 2 h v (R) =clip(r + v, 0, 1) 3 H v (R) :=2I hv (R) Approximate full precision activations using binary bases A 1,, A N = H v1 (R),,H vn (R) R 1 A N A N

11 Approximate full precision activations by shift Construct binary activation bases by shift 1 R = ReLU(Input) 2 h v (R) =clip(r + v, 0, 1) 3 H v (R) :=2I hv (R) Approximate full precision activations using binary bases A 1,, A N = H v1 (R),,H vn (R) R 1 A N A N

12 Approximate full precision activations by shift Construct binary activation bases by shift 1 R = ReLU(Input) 2 h v (R) =clip(r + v, 0, 1) 3 H v (R) :=2I hv (R) Approximate full precision activations using binary bases A 1,, A N = H v1 (R),,H vn (R) R 1 A N A N

13 Parallel & multiple binary convolution Conv(W, R)! MX NX Conv m B m, na n NX = m=1 nconv MX n=1 m B m, A n! = n=1 m=1 m=1 n=1 sum-sum operation can be parallel MX NX m n Conv (B m, A n ) Binary Conv: x y = POPCOUNT(x XNOR y),x i,y i 2 { 1, +1}, 8i Advantages 1. Bitwise operations 2. More bases, better approximation 3. Parallel computation (sum-sum) is hardware friendly

14 Result Model on ImageNet Benchmark Full-Precision ResNet-18 [full-precision weights and activations] BWN [full-precision activation] [8] Rastegari et al. (2016) DoReFa-Net [1-bit weight and 4-bit activation] [4] Zhou et al. (2016) XNOR-Net [binary weight and activation] [8] Rastegari et al. (2016) BNN [binary weight and activation] [7] Courbariaux et al. (2016) weight (bit) activation (bit) Accuracy (Top-1) Accuracy Loss % % 7.5% % 10.1% % 18.1% % 27.1% Ours [3 weight bases, 3 activation bases] % 5.4% Ours [5 weight bases, 5 activation bases] % 4.3% Full-Precision ResNet-34 [full-precision weights and activations] % Ours [5 weight bases, 5 activation bases] % 4.9%

15 Future work 1.Applicability to other tasks, e.g., object detection, parsing, face recognition, speech recognition, etc. 2.Hardware acceleration on mobile 3.Software acceleration on cloud Paper: #261 Poster: Pacific Ballroom #101 Contact: Thank You!

TYPES OF MODEL COMPRESSION. Soham Saha, MS by Research in CSE, CVIT, IIIT Hyderabad

TYPES OF MODEL COMPRESSION. Soham Saha, MS by Research in CSE, CVIT, IIIT Hyderabad TYPES OF MODEL COMPRESSION Soham Saha, MS by Research in CSE, CVIT, IIIT Hyderabad 1. Pruning 2. Quantization 3. Architectural Modifications PRUNING WHY PRUNING? Deep Neural Networks have redundant parameters.

More information

Neural Network Approximation. Low rank, Sparsity, and Quantization Oct. 2017

Neural Network Approximation. Low rank, Sparsity, and Quantization Oct. 2017 Neural Network Approximation Low rank, Sparsity, and Quantization zsc@megvii.com Oct. 2017 Motivation Faster Inference Faster Training Latency critical scenarios VR/AR, UGV/UAV Saves time and energy Higher

More information

Multi-Precision Quantized Neural Networks via Encoding Decomposition of { 1, +1}

Multi-Precision Quantized Neural Networks via Encoding Decomposition of { 1, +1} Multi-Precision Quantized Neural Networks via Encoding Decomposition of {, +} Qigong Sun, Fanhua Shang, Kang Yang, Xiufang Li, Yan Ren, Licheng Jiao Key Laboratory of Intelligent Perception and Image Understanding

More information

Quantisation. Efficient implementation of convolutional neural networks. Philip Leong. Computer Engineering Lab The University of Sydney

Quantisation. Efficient implementation of convolutional neural networks. Philip Leong. Computer Engineering Lab The University of Sydney 1/51 Quantisation Efficient implementation of convolutional neural networks Philip Leong Computer Engineering Lab The University of Sydney July 2018 / PAPAA Workshop Australia 2/51 3/51 Outline 1 Introduction

More information

Differentiable Fine-grained Quantization for Deep Neural Network Compression

Differentiable Fine-grained Quantization for Deep Neural Network Compression Differentiable Fine-grained Quantization for Deep Neural Network Compression Hsin-Pai Cheng hc218@duke.edu Yuanjun Huang University of Science and Technology of China Anhui, China yjhuang@mail.ustc.edu.cn

More information

Multiscale methods for neural image processing. Sohil Shah, Pallabi Ghosh, Larry S. Davis and Tom Goldstein Hao Li, Soham De, Zheng Xu, Hanan Samet

Multiscale methods for neural image processing. Sohil Shah, Pallabi Ghosh, Larry S. Davis and Tom Goldstein Hao Li, Soham De, Zheng Xu, Hanan Samet Multiscale methods for neural image processing Sohil Shah, Pallabi Ghosh, Larry S. Davis and Tom Goldstein Hao Li, Soham De, Zheng Xu, Hanan Samet A TALK IN TWO ACTS Part I: Stacked U-nets The globalization

More information

arxiv: v1 [cs.lg] 10 Nov 2017

arxiv: v1 [cs.lg] 10 Nov 2017 Quantized Memory-Augmented Neural Networks Seongsik Park, Seijoon Kim, Seil Lee, Ho Bae 2, and Sungroh Yoon,2 Department of Electrical and Computer Engineering, Seoul National University, Seoul 8826, Korea

More information

arxiv: v1 [cs.cv] 25 May 2018

arxiv: v1 [cs.cv] 25 May 2018 Heterogeneous Bitwidth Binarization in Convolutional Neural Networks arxiv:1805.10368v1 [cs.cv] 25 May 2018 Josh Fromm Department of Electrical Engineering University of Washington Seattle, WA 98195 jwfromm@uw.edu

More information

LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks. Microsoft Research

LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks. Microsoft Research LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua Microsoft Research zdqzeros@gmail.com jiaoyan@microsoft.com

More information

Quantization of Fully Convolutional Networks for Accurate Biomedical Image Segmentation

Quantization of Fully Convolutional Networks for Accurate Biomedical Image Segmentation Quantization of Fully Convolutional Networks for Accurate Biomedical Image Segmentation Xiaowei Xu 1,, Qing Lu 1, Yu Hu, Lin Yang 1, Sharon Hu 1, Danny Chen 1, Yiyu Shi 1 1 Univerity of Notre Dame Huazhong

More information

Two-Step Quantization for Low-bit Neural Networks

Two-Step Quantization for Low-bit Neural Networks Two-Step Quantization for Low-bit Neural Networks Peisong Wang 1,2, Qinghao Hu 1,2, Yifan Zhang 1,2, Chunjie Zhang 1,2, Yang Liu 4, and Jian Cheng 1,2,3 1 Institute of Automation, Chinese Academy of Sciences,

More information

Binary Deep Learning. Presented by Roey Nagar and Kostya Berestizshevsky

Binary Deep Learning. Presented by Roey Nagar and Kostya Berestizshevsky Binary Deep Learning Presented by Roey Nagar and Kostya Berestizshevsky Deep Learning Seminar, School of Electrical Engineering, Tel Aviv University January 22 nd 2017 Lecture Outline Motivation and existing

More information

Scalable Methods for 8-bit Training of Neural Networks

Scalable Methods for 8-bit Training of Neural Networks Scalable Methods for 8-bit Training of Neural Networks Ron Banner 1, Itay Hubara 2, Elad Hoffer 2, Daniel Soudry 2 {itayhubara, elad.hoffer, daniel.soudry}@gmail.com {ron.banner}@intel.com (1) Intel -

More information

Layer 1. Layer L. difference between the two precisions (B A and B W ) is chosen to balance the sum in (1), as follows: B A B W = round log 2 (2)

Layer 1. Layer L. difference between the two precisions (B A and B W ) is chosen to balance the sum in (1), as follows: B A B W = round log 2 (2) AN ANALYTICAL METHOD TO DETERMINE MINIMUM PER-LAYER PRECISION OF DEEP NEURAL NETWORKS Charbel Sakr Naresh Shanbhag Dept. of Electrical Computer Engineering, University of Illinois at Urbana Champaign ABSTRACT

More information

arxiv: v4 [cs.cv] 2 Aug 2016

arxiv: v4 [cs.cv] 2 Aug 2016 arxiv:1603.05279v4 [cs.cv] 2 Aug 2016 XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, Ali Farhadi Allen Institute for AI,

More information

Stochastic Layer-Wise Precision in Deep Neural Networks

Stochastic Layer-Wise Precision in Deep Neural Networks Stochastic Layer-Wise Precision in Deep Neural Networks Griffin Lacey NVIDIA Graham W. Taylor University of Guelph Vector Institute for Artificial Intelligence Canadian Institute for Advanced Research

More information

Frequency-Domain Dynamic Pruning for Convolutional Neural Networks

Frequency-Domain Dynamic Pruning for Convolutional Neural Networks Frequency-Domain Dynamic Pruning for Convolutional Neural Networks Zhenhua Liu 1, Jizheng Xu 2, Xiulian Peng 2, Ruiqin Xiong 1 1 Institute of Digital Media, School of Electronic Engineering and Computer

More information

σ(x) = clip( x m + α 2, 0, α) = min(max( x m + α 2, 0), α) (1)

σ(x) = clip( x m + α 2, 0, α) = min(max( x m + α 2, 0), α) (1) TRUE GRADIENT-BASED TRAINING OF DEEP BINARY ACTIVATED NEURAL NETWORKS VIA CONTINUOUS BINARIZATION Charbel Sakr, Jungwook Choi, Zhuo Wang, Kailash Gopalakrishnan, Naresh Shanbhag Dept. of Electrical and

More information

Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations

Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations Journal of Machine Learning Research 18 (2018) 1-30 Submitted 9/16; Revised 4/17; Published 4/18 Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations Itay Hubara*

More information

Weighted-Entropy-based Quantization for Deep Neural Networks

Weighted-Entropy-based Quantization for Deep Neural Networks Weighted-Entropy-based Quantization for Deep Neural Networks Eunhyeok Park, Junwhan Ahn, and Sungjoo Yoo canusglow@gmail.com, junwhan@snu.ac.kr, sungjoo.yoo@gmail.com Seoul National University Computing

More information

Deep Learning with Low Precision by Half-wave Gaussian Quantization

Deep Learning with Low Precision by Half-wave Gaussian Quantization Deep Learning with Low Precision by Half-wave Gaussian Quantization Zhaowei Cai UC San Diego zwcai@ucsd.edu Xiaodong He Microsoft Research Redmond xiaohe@microsoft.com Jian Sun Megvii Inc. sunjian@megvii.com

More information

<Special Topics in VLSI> Learning for Deep Neural Networks (Back-propagation)

<Special Topics in VLSI> Learning for Deep Neural Networks (Back-propagation) Learning for Deep Neural Networks (Back-propagation) Outline Summary of Previous Standford Lecture Universal Approximation Theorem Inference vs Training Gradient Descent Back-Propagation

More information

arxiv: v2 [cs.lg] 3 Dec 2018

arxiv: v2 [cs.lg] 3 Dec 2018 Binary Ensemble Neural Network: More Bits per Network or More Networks per Bit? arxiv:1806.07550v2 [cs.lg] 3 Dec 2018 Shilin Zhu UC San Diego La Jolla, CA 92093 shz338@eng.ucsd.edu Abstract Binary neural

More information

arxiv: v2 [cs.cv] 20 Aug 2018

arxiv: v2 [cs.cv] 20 Aug 2018 Joint Training of Low-Precision Neural Network with Quantization Interval Parameters arxiv:1808.05779v2 [cs.cv] 20 Aug 2018 Sangil Jung sang-il.jung Youngjun Kwak yjk.kwak Changyong Son cyson Jae-Joon

More information

arxiv: v2 [cs.cv] 17 Nov 2017

arxiv: v2 [cs.cv] 17 Nov 2017 Towards Effective Low-bitwidth Convolutional Neural Networks Bohan Zhuang, Chunhua Shen, Mingkui Tan, Lingqiao Liu, Ian Reid arxiv:1711.00205v2 [cs.cv] 17 Nov 2017 Abstract This paper tackles the problem

More information

arxiv: v1 [cs.cv] 16 Mar 2016

arxiv: v1 [cs.cv] 16 Mar 2016 arxiv:1603.05279v1 [cs.cv] 16 Mar 2016 XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, Ali Farhadi Allen Institute for AI,

More information

DNN FEATURE MAP COMPRESSION USING LEARNED REPRESENTATION OVER GF(2)

DNN FEATURE MAP COMPRESSION USING LEARNED REPRESENTATION OVER GF(2) DNN FEATURE MAP COMPRESSION USING LEARNED REPRESENTATION OVER GF(2) Anonymous authors Paper under double-blind review ABSTRACT In this paper, we introduce a method to compress intermediate feature maps

More information

On the Complexity of Neural-Network-Learned Functions

On the Complexity of Neural-Network-Learned Functions On the Complexity of Neural-Network-Learned Functions Caren Marzban 1, Raju Viswanathan 2 1 Applied Physics Lab, and Dept of Statistics, University of Washington, Seattle, WA 98195 2 Cyberon LLC, 3073

More information

arxiv: v1 [stat.ml] 18 Jan 2019

arxiv: v1 [stat.ml] 18 Jan 2019 Foothill: A Quasiconvex Regularization Function Mouloud Belbahri, Eyyüb Sari, Sajad Darabi, Vahid Partovi Nia arxiv:90.0644v [stat.ml] 8 Jan 09 Huawei Technologies Co., Ltd. Montreal Research Center, Canada

More information

Auto-balanced Filter Pruning for Efficient Convolutional Neural Networks

Auto-balanced Filter Pruning for Efficient Convolutional Neural Networks Auto-balanced Filter Pruning for Efficient Convolutional Neural Networks Xiaohan Ding, 1 Guiguang Ding, 1 Jungong Han, 2 Sheng Tang 3 1 School of Software, Tsinghua University, Beijing 100084, China 2

More information

arxiv: v1 [cs.cv] 3 Feb 2017

arxiv: v1 [cs.cv] 3 Feb 2017 Deep Learning with Low Precision by Half-wave Gaussian Quantization Zhaowei Cai UC San Diego zwcai@ucsd.edu Xiaodong He Microsoft Research Redmond xiaohe@microsoft.com Jian Sun Megvii Inc. sunjian@megvii.com

More information

arxiv: v1 [cs.ne] 20 Apr 2018

arxiv: v1 [cs.ne] 20 Apr 2018 arxiv:1804.07802v1 [cs.ne] 20 Apr 2018 Value-aware Quantization for Training and Inference of Neural Networks Eunhyeok Park 1, Sungjoo Yoo 1, Peter Vajda 2 1 Department of Computer Science and Engineering

More information

Deep Neural Network Compression with Single and Multiple Level Quantization

Deep Neural Network Compression with Single and Multiple Level Quantization The Thirty-Second AAAI Conference on Artificial Intelligence (AAAI-18) Deep Neural Network Compression with Single and Multiple Level Quantization Yuhui Xu, 1 Yongzhuang Wang, 1 Aojun Zhou, 2 Weiyao Lin,

More information

Quantization of Fully Convolutional Networks for Accurate Biomedical Image Segmentation

Quantization of Fully Convolutional Networks for Accurate Biomedical Image Segmentation Quantization of Fully Convolutional Networks for Accurate Biomedical Image Segmentation Xiaowei Xu 1,2, Qing Lu 2, Lin Yang 2, Sharon Hu 2, Danny Chen 2, Yu Hu 1, Yiyu Shi 2 1 Huazhong University of Science

More information

Performance Guaranteed Network Acceleration via High-Order Residual Quantization

Performance Guaranteed Network Acceleration via High-Order Residual Quantization Performance Guaranteed Network Acceleration via High-Order Residual Quantization Zefan Li, Bingbing Ni, Wenjun Zhang, Xiaokang Yang, Wen Gao 2 Shanghai Jiao Tong University, 2 Peking University {Leezf,

More information

Efficient DNN Neuron Pruning by Minimizing Layer-wise Nonlinear Reconstruction Error

Efficient DNN Neuron Pruning by Minimizing Layer-wise Nonlinear Reconstruction Error Efficient DNN Neuron Pruning by Minimizing Layer-wise Nonlinear Reconstruction Error Chunhui Jiang, Guiying Li, Chao Qian, Ke Tang Anhui Province Key Lab of Big Data Analysis and Application, University

More information

Towards Effective Low-bitwidth Convolutional Neural Networks

Towards Effective Low-bitwidth Convolutional Neural Networks Towards Effective Low-bitwidth Convolutional Neural Networks Bohan Zhuang 1,2, Chunhua Shen 1,2, Mingkui Tan 3, Lingqiao Liu 1, Ian Reid 1,2 1 The University of Adelaide, Australia, 2 Australian Centre

More information

Compressing deep neural networks

Compressing deep neural networks From Data to Decisions - M.Sc. Data Science Compressing deep neural networks Challenges and theoretical foundations Presenter: Simone Scardapane University of Exeter, UK Table of contents Introduction

More information

arxiv: v2 [cs.ne] 23 Nov 2017

arxiv: v2 [cs.ne] 23 Nov 2017 Minimum Energy Quantized Neural Networks Bert Moons +, Koen Goetschalckx +, Nick Van Berckelaer* and Marian Verhelst + Department of Electrical Engineering* + - ESAT/MICAS +, KU Leuven, Leuven, Belgium

More information

Rethinking Binary Neural Network for Accurate Image Classification and Semantic Segmentation

Rethinking Binary Neural Network for Accurate Image Classification and Semantic Segmentation Rethinking nary Neural Network for Accurate Image Classification and Semantic Segmentation Bohan Zhuang 1, Chunhua Shen 1, Mingkui Tan 2, Lingqiao Liu 1, and Ian Reid 1 1 The University of Adelaide, Australia;

More information

arxiv: v1 [cs.cv] 4 Apr 2019

arxiv: v1 [cs.cv] 4 Apr 2019 Regularizing Activation Distribution for Training Binarized Deep Networks Ruizhou Ding, Ting-Wu Chin, Zeye Liu, Diana Marculescu Carnegie Mellon University {rding, tingwuc, zeyel, dianam}@andrew.cmu.edu

More information

arxiv: v2 [cs.cv] 6 Dec 2018

arxiv: v2 [cs.cv] 6 Dec 2018 HAQ: Hardware-Aware Automated Quantization Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han {kuanwang, zhijian, yujunlin, jilin, songhan}@mit.edu Massachusetts Institute of Technology arxiv:1811.0888v

More information

arxiv: v2 [cs.cv] 19 Sep 2017

arxiv: v2 [cs.cv] 19 Sep 2017 Learning Convolutional Networks for Content-weighted Image Compression arxiv:1703.10553v2 [cs.cv] 19 Sep 2017 Mu Li Hong Kong Polytechnic University csmuli@comp.polyu.edu.hk Shuhang Gu Hong Kong Polytechnic

More information

Value-aware Quantization for Training and Inference of Neural Networks

Value-aware Quantization for Training and Inference of Neural Networks Value-aware Quantization for Training and Inference of Neural Networks Eunhyeok Park 1, Sungjoo Yoo 1, and Peter Vajda 2 1 Department of Computer Science and Engineering Seoul National University {eunhyeok.park,sungjoo.yoo}@gmail.com

More information

Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference

Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference Benoit Jacob Skirmantas Kligys Bo Chen Menglong Zhu Matthew Tang Andrew Howard Hartwig Adam Dmitry Kalenichenko

More information

Designing Energy-Efficient Convolutional Neural Networks using Energy-Aware Pruning

Designing Energy-Efficient Convolutional Neural Networks using Energy-Aware Pruning Designing Energy-Efficient Convolutional Neural Networks using Energy-Aware Pruning Tien-Ju Yang, Yu-Hsin Chen, Vivienne Sze Massachusetts Institute of Technology {tjy, yhchen, sze}@mit.edu Abstract Deep

More information

arxiv: v3 [cs.lg] 2 Jun 2016

arxiv: v3 [cs.lg] 2 Jun 2016 Darryl D. Lin Qualcomm Research, San Diego, CA 922, USA Sachin S. Talathi Qualcomm Research, San Diego, CA 922, USA DARRYL.DLIN@GMAIL.COM TALATHI@GMAIL.COM arxiv:5.06393v3 [cs.lg] 2 Jun 206 V. Sreekanth

More information

arxiv: v1 [cs.cv] 3 Aug 2017

arxiv: v1 [cs.cv] 3 Aug 2017 DONG ET AL.: STOCHASTIC QUANTIZATION 1 arxiv:1708.01001v1 [cs.cv] 3 Aug 2017 Learning Accurate Low-Bit Deep Neural Networks with Stochastic Quantization Yinpeng Dong 1 dyp17@mails.tsinghua.edu.cn Renkun

More information

Towards a Deeper Understanding of Training Quantized Neural Networks

Towards a Deeper Understanding of Training Quantized Neural Networks Hao Li * 1 Soham De * 1 Zheng Xu 1 Christoph Studer Hanan Samet 1 Tom Goldstein 1 Abstract Training neural networks with coarsely quantized weights is a key step towards learning on embedded platforms

More information

arxiv: v1 [cs.lg] 10 Jan 2019

arxiv: v1 [cs.lg] 10 Jan 2019 Quantized Epoch-SGD for Communication-Efficient Distributed Learning arxiv:1901.0300v1 [cs.lg] 10 Jan 2019 Shen-Yi Zhao Hao Gao Wu-Jun Li Department of Computer Science and Technology Nanjing University,

More information

Deep learning on 3D geometries. Hope Yao Design Informatics Lab Department of Mechanical and Aerospace Engineering

Deep learning on 3D geometries. Hope Yao Design Informatics Lab Department of Mechanical and Aerospace Engineering Deep learning on 3D geometries Hope Yao Design Informatics Lab Department of Mechanical and Aerospace Engineering Overview Background Methods Numerical Result Future improvements Conclusion Background

More information

Fixed-point Factorized Networks

Fixed-point Factorized Networks Fixed-point Factorized Networks Peisong Wang 1,2 and Jian Cheng 1,2,3 1 Institute of Automation, Chinese Academy of Sciences 2 University of Chinese Academy of Sciences 3 Center for Excellence in Brain

More information

FxpNet: Training deep convolutional neural network in fixed-point representation

FxpNet: Training deep convolutional neural network in fixed-point representation FxpNet: Training deep convolutional neural networ in fixed-point representation Xi Chen Department of Computer Science and Technology Tsinghua University 100084, Beijing, China aaron.xichen@gmail.com Xiaolin

More information

TBN: Convolutional Neural Network with Ternary Inputs and Binary Weights

TBN: Convolutional Neural Network with Ternary Inputs and Binary Weights TBN: Convolutional Neural Network with Ternary Inputs and Binary Weights Diwen Wan 1,2, Fumin Shen 1, Li Liu 2, Fan Zhu 2, Jie Qin 3, Ling Shao 2, and Heng Tao Shen 1 1 Center for Future Media and School

More information

Deep Learning Year in Review 2016: Computer Vision Perspective

Deep Learning Year in Review 2016: Computer Vision Perspective Deep Learning Year in Review 2016: Computer Vision Perspective Alex Kalinin, PhD Candidate Bioinformatics @ UMich alxndrkalinin@gmail.com @alxndrkalinin Architectures Summary of CNN architecture development

More information

Analytical Guarantees on Numerical Precision of Deep Neural Networks

Analytical Guarantees on Numerical Precision of Deep Neural Networks Charbel Sakr Yongjune Kim Naresh Shanbhag Abstract The acclaimed successes of neural networks often overshadow their tremendous complexity. We focus on numerical precision - a key parameter defining the

More information

Binary Convolutional Neural Network on RRAM

Binary Convolutional Neural Network on RRAM Binary Convolutional Neural Network on RRAM Tianqi Tang, Lixue Xia, Boxun Li, Yu Wang, Huazhong Yang Dept. of E.E, Tsinghua National Laboratory for Information Science and Technology (TNList) Tsinghua

More information

Accelerating Convolutional Neural Networks by Group-wise 2D-filter Pruning

Accelerating Convolutional Neural Networks by Group-wise 2D-filter Pruning Accelerating Convolutional Neural Networks by Group-wise D-filter Pruning Niange Yu Department of Computer Sicence and Technology Tsinghua University Beijing 0008, China yng@mails.tsinghua.edu.cn Shi Qiu

More information

Convolutional Neural Network Architecture

Convolutional Neural Network Architecture Convolutional Neural Network Architecture Zhisheng Zhong Feburary 2nd, 2018 Zhisheng Zhong Convolutional Neural Network Architecture Feburary 2nd, 2018 1 / 55 Outline 1 Introduction of Convolution Motivation

More information

FINN-L: Library Extensions and Design Trade-off Analysis for Variable Precision LSTM Networks on FPGAs

FINN-L: Library Extensions and Design Trade-off Analysis for Variable Precision LSTM Networks on FPGAs FINN-L: Library Extensions and Design Trade-off Analysis for Variable Precision LSTM Networks on FPGAs Vladimir Rybalkin, Muhammad Mohsin Ghaffar and Norbert Wehn Microelectronic Systems Design Research

More information

A MAIN/SUBSIDIARY NETWORK FRAMEWORK FOR SIMPLIFYING BINARY NEURAL NETWORKS

A MAIN/SUBSIDIARY NETWORK FRAMEWORK FOR SIMPLIFYING BINARY NEURAL NETWORKS Under review as a conference paper at ICLR 209 A MAIN/SUBSIDIARY NETWORK FRAMEWORK FOR SIMPLIFYING BINARY NEURAL NETWORKS Anonymous authors Paper under double-blind review ABSTRACT To reduce memory footprint

More information

arxiv: v2 [cs.cv] 7 Dec 2017

arxiv: v2 [cs.cv] 7 Dec 2017 ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices Xiangyu Zhang Xinyu Zhou Mengxiao Lin Jian Sun Megvii Inc (Face++) {zhangxiangyu,zxy,linmengxiao,sunjian}@megvii.com arxiv:1707.01083v2

More information

Recent Advances in Efficient Computation of Deep Convolutional Neural Networks

Recent Advances in Efficient Computation of Deep Convolutional Neural Networks Recent Advances in Efficient Computation of Deep Convolutional Neural Networks however, is that the computational complexity as well as the storage requirements of these DNNs has also increased drastically

More information

Discrimination-aware Channel Pruning for Deep Neural Networks

Discrimination-aware Channel Pruning for Deep Neural Networks Discrimination-aware Channel Pruning for Deep Neural Networks Zhuangwei Zhuang 1, Mingkui Tan 1, Bohan Zhuang 2, Jing Liu 1, Yong Guo 1, Qingyao Wu 1, Junzhou Huang 3,4, Jinhui Zhu 1 1 South China University

More information

TRAINING AND INFERENCE WITH INTEGERS IN DEEP NEURAL NETWORKS

TRAINING AND INFERENCE WITH INTEGERS IN DEEP NEURAL NETWORKS Published as a conference paper at ICLR 218 TRAINING AND INFERENCE WITH INTEGERS IN DEEP NEURAL NETWORKS Shuang Wu 1, Guoqi Li 1, Feng Chen 2, Luping Shi 1 1 Department of Precision Instrument 2 Department

More information

Analytical Guarantees on Numerical Precision of Deep Neural Networks

Analytical Guarantees on Numerical Precision of Deep Neural Networks Charbel Sakr Yongjune Kim Naresh Shanbhag Abstract The acclaimed successes of neural networks often overshadow their tremendous complexity. We focus on numerical precision - a key parameter defining the

More information

CondenseNet: An Efficient DenseNet using Learned Group Convolutions

CondenseNet: An Efficient DenseNet using Learned Group Convolutions CondenseNet: An Efficient DenseNet using Learned Group s Gao Huang Cornell University gh@cornell.edu Shichen Liu Tsinghua University liushichen@gmail.com Kilian Q. Weinberger Cornell University kqw@cornell.edu

More information

arxiv: v1 [cs.cv] 31 Jan 2018

arxiv: v1 [cs.cv] 31 Jan 2018 Recovering from Random Pruning: On the Plasticity of Deep Convolutional Neural Networks Deepak Mittal Shweta Bhardwaj Mitesh M. Khapra Balaraman Ravindran arxiv:1801.10447v1 [cs.cv] 31 Jan 2018 Department

More information

Improved Bayesian Compression

Improved Bayesian Compression Improved Bayesian Compression Marco Federici University of Amsterdam marco.federici@student.uva.nl Karen Ullrich University of Amsterdam karen.ullrich@uva.nl Max Welling University of Amsterdam Canadian

More information

ADAPTIVE QUANTIZATION OF NEURAL NETWORKS

ADAPTIVE QUANTIZATION OF NEURAL NETWORKS ADAPTIVE QUANTIZATION OF NEURAL NETWORKS Soroosh Khoram Department of Electrical and Computer Engineering University of Wisconsin - Madison khoram@wisc.edu Jing Li Department of Electrical and Computer

More information

arxiv: v2 [cs.cv] 17 Jul 2018 ABSTRACT

arxiv: v2 [cs.cv] 17 Jul 2018 ABSTRACT PACT: PARAMETERIZED CLIPPING ACTIVATION FOR QUANTIZED NEURAL NETWORKS Jungwook Choi 1, Zhuo Wang 2, Swagath Venkataramani 2, Pierce I-Jen Chuang 1, Vijayalakshmi Srinivasan 1, Kailash Gopalakrishnan 1

More information

Global Optimality in Matrix and Tensor Factorization, Deep Learning & Beyond

Global Optimality in Matrix and Tensor Factorization, Deep Learning & Beyond Global Optimality in Matrix and Tensor Factorization, Deep Learning & Beyond Ben Haeffele and René Vidal Center for Imaging Science Mathematical Institute for Data Science Johns Hopkins University This

More information

Efficient Deep Learning Inference based on Model Compression

Efficient Deep Learning Inference based on Model Compression Efficient Deep Learning Inference based on Model Compression Qing Zhang, Mengru Zhang, Mengdi Wang, Wanchen Sui, Chen Meng, Jun Yang Alibaba Group {sensi.zq, mengru.zmr, didou.wmd, wanchen.swc, mc119496,

More information

CondenseNet: An Efficient DenseNet using Learned Group Convolutions

CondenseNet: An Efficient DenseNet using Learned Group Convolutions CondenseNet: An Efficient DenseNet using Learned Group s Gao Huang Cornell University gh@cornell.edu Shichen Liu Tsinghua University liushichen@gmail.com Kilian Q. Weinberger Cornell University kqw@cornell.edu

More information

arxiv: v3 [cs.ne] 2 Feb 2018

arxiv: v3 [cs.ne] 2 Feb 2018 DOREFA-NET: TRAINING LOW BITWIDTH CONVOLU- TIONAL NEURAL NETWORKS WITH LOW BITWIDTH GRADIENTS Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, Yuheng Zou Megvii Inc. {zsc, wyx, nzk, zxy, wenhe, zouyuheng}@megvii.com

More information

arxiv: v2 [cs.cv] 30 Oct 2018

arxiv: v2 [cs.cv] 30 Oct 2018 Discrimination-aware Channel Pruning for Deep Neural Networks arxiv:1810.11809v2 [cs.cv] 30 Oct 2018 Zhuangwei Zhuang 1, Mingkui Tan 1, Bohan Zhuang 2, Jing Liu 1, Yong Guo 1, Qingyao Wu 1, Junzhou Huang

More information

arxiv: v1 [cs.lg] 19 May 2017

arxiv: v1 [cs.lg] 19 May 2017 The High-Dimensional Geometry of Binary Neural Networks arxiv:175.7199v1 [cs.lg] 19 May 217 Alexander G. Anderson Redwood Center for Theoretical Neuroscience University of California, Berkeley aga@berkeley.edu

More information

PRUNING CONVOLUTIONAL NEURAL NETWORKS. Pavlo Molchanov Stephen Tyree Tero Karras Timo Aila Jan Kautz

PRUNING CONVOLUTIONAL NEURAL NETWORKS. Pavlo Molchanov Stephen Tyree Tero Karras Timo Aila Jan Kautz PRUNING CONVOLUTIONAL NEURAL NETWORKS Pavlo Molchanov Stephen Tyree Tero Karras Timo Aila Jan Kautz 2017 WHY WE CAN PRUNE CNNS? 2 WHY WE CAN PRUNE CNNS? Optimization failures : Some neurons are "dead":

More information

arxiv: v1 [cs.cv] 11 Sep 2018

arxiv: v1 [cs.cv] 11 Sep 2018 Discovering Low-Precision Networks Close to Full-Precision Networks for Efficient Embedded Inference Jeffrey L. McKinstry jlmckins@us.ibm.com Steven K. Esser sesser@us.ibm.com Rathinakumar Appuswamy rappusw@us.ibm.com

More information

CSC321 Lecture 16: ResNets and Attention

CSC321 Lecture 16: ResNets and Attention CSC321 Lecture 16: ResNets and Attention Roger Grosse Roger Grosse CSC321 Lecture 16: ResNets and Attention 1 / 24 Overview Two topics for today: Topic 1: Deep Residual Networks (ResNets) This is the state-of-the

More information

arxiv: v1 [cs.cv] 6 Dec 2018

arxiv: v1 [cs.cv] 6 Dec 2018 Trained Rank Pruning for Efficient Deep Neural Networks Yuhui Xu 1, Yuxi Li 1, Shuai Zhang 2, Wei Wen 3, Botao Wang 2, Yingyong Qi 2, Yiran Chen 3, Weiyao Lin 1 and Hongkai Xiong 1 arxiv:1812.02402v1 [cs.cv]

More information

Perturbative Neural Networks

Perturbative Neural Networks Perturbative Neural Networks Felix Juefei-Xu Carnegie Mellon University felixu@cmu.edu Vishnu Naresh Boddeti Michigan State University vishnu@msu.edu Marios Savvides Carnegie Mellon University msavvid@ri.cmu.edu

More information

SYQ: Learning Symmetric Quantization For Efficient Deep Neural Networks

SYQ: Learning Symmetric Quantization For Efficient Deep Neural Networks SYQ: Learning Symmetric Quantization For Efficient Deep Neural Networks Julian Faraone* Nicholas Fraser # Michaela Blott # Philip H.W. Leong* The University of Sydney* Xilinx Research Labs # (julian.faraone,

More information

arxiv: v1 [stat.ml] 15 Dec 2017

arxiv: v1 [stat.ml] 15 Dec 2017 BT-Nets: Simplifying Deep Neural Networks via Block Term Decomposition Guangxi Li 1, Jinmian Ye 1, Haiqin Yang 2, Di Chen 1, Shuicheng Yan 3, Zenglin Xu 1, 1 SMILE Lab, School of Comp. Sci. and Eng., Univ.

More information

Sparse Regularized Deep Neural Networks For Efficient Embedded Learning

Sparse Regularized Deep Neural Networks For Efficient Embedded Learning Sparse Regularized Deep Neural Networks For Efficient Embedded Learning Anonymous authors Paper under double-blind review Abstract Deep learning is becoming more widespread in its application due to its

More information

Tasks ADAS. Self Driving. Non-machine Learning. Traditional MLP. Machine-Learning based method. Supervised CNN. Methods. Deep-Learning based

Tasks ADAS. Self Driving. Non-machine Learning. Traditional MLP. Machine-Learning based method. Supervised CNN. Methods. Deep-Learning based UNDERSTANDING CNN ADAS Tasks Self Driving Localizati on Perception Planning/ Control Driver state Vehicle Diagnosis Smart factory Methods Traditional Deep-Learning based Non-machine Learning Machine-Learning

More information

Large-Scale FPGA implementations of Machine Learning Algorithms

Large-Scale FPGA implementations of Machine Learning Algorithms Large-Scale FPGA implementations of Machine Learning Algorithms Philip Leong ( ) Computer Engineering Laboratory School of Electrical and Information Engineering, The University of Sydney Computer Engineering

More information

Sajid Anwar, Kyuyeon Hwang and Wonyong Sung

Sajid Anwar, Kyuyeon Hwang and Wonyong Sung Sajid Anwar, Kyuyeon Hwang and Wonyong Sung Department of Electrical and Computer Engineering Seoul National University Seoul, 08826 Korea Email: sajid@dsp.snu.ac.kr, khwang@dsp.snu.ac.kr, wysung@snu.ac.kr

More information

Attention Based Pruning for Shift Networks

Attention Based Pruning for Shift Networks Attention Based Pruning for Shift Networks Ghouthi Boukli Hacene, Ghouthi Hacene, Carlos Lassance, Vincent Gripon To cite this version: Ghouthi Boukli Hacene, Ghouthi Hacene, Carlos Lassance, Vincent Gripon.

More information

An Analytical Method to Determine Minimum Per-Layer Precision of Deep Neural Networks

An Analytical Method to Determine Minimum Per-Layer Precision of Deep Neural Networks An Analytical Method to Determine Minimum Per-Layer Precision of Deep Neural Networks Charbel Sakr, Naresh Shanbhag Department of Electrical and Computer Engineering University of Illinois at Urbana-Champaign

More information

RAGAV VENKATESAN VIJETHA GATUPALLI BAOXIN LI NEURAL DATASET GENERALITY

RAGAV VENKATESAN VIJETHA GATUPALLI BAOXIN LI NEURAL DATASET GENERALITY RAGAV VENKATESAN VIJETHA GATUPALLI BAOXIN LI NEURAL DATASET GENERALITY SIFT HOG ALL ABOUT THE FEATURES DAISY GABOR AlexNet GoogleNet CONVOLUTIONAL NEURAL NETWORKS VGG-19 ResNet FEATURES COMES FROM DATA

More information

arxiv: v1 [cs.lg] 15 Dec 2017

arxiv: v1 [cs.lg] 15 Dec 2017 Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference arxiv:1712.05877v1 [cs.lg] 15 Dec 2017 Benoit Jacob Skirmantas Kligys Bo Chen Menglong Zhu Matthew Tang Andrew

More information

arxiv: v1 [cs.ne] 20 Jun 2016

arxiv: v1 [cs.ne] 20 Jun 2016 DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients Shuchang Zhou, Zekun Ni, Xinyu Zhou, He Wen, Yuxin Wu, Yuheng Zou Megvii Inc. {shuchang.zhou,mike.zekun}@gmail.com

More information

Very Deep Residual Networks with Maxout for Plant Identification in the Wild Milan Šulc, Dmytro Mishkin, Jiří Matas

Very Deep Residual Networks with Maxout for Plant Identification in the Wild Milan Šulc, Dmytro Mishkin, Jiří Matas Very Deep Residual Networks with Maxout for Plant Identification in the Wild Milan Šulc, Dmytro Mishkin, Jiří Matas Center for Machine Perception Department of Cybernetics Faculty of Electrical Engineering

More information

Normalization Techniques in Training of Deep Neural Networks

Normalization Techniques in Training of Deep Neural Networks Normalization Techniques in Training of Deep Neural Networks Lei Huang ( 黄雷 ) State Key Laboratory of Software Development Environment, Beihang University Mail:huanglei@nlsde.buaa.edu.cn August 17 th,

More information

ACIQ: ANALYTICAL CLIPPING FOR INTEGER QUAN-

ACIQ: ANALYTICAL CLIPPING FOR INTEGER QUAN- ACIQ: ANALYTICAL CLIPPING FOR INTEGER QUAN- TIZATION OF NEURAL NETWORKS Anonymous authors Paper under double-blind review ABSTRACT Unlike traditional approaches that focus on the quantization at the network

More information

Faster Machine Learning via Low-Precision Communication & Computation. Dan Alistarh (IST Austria & ETH Zurich), Hantian Zhang (ETH Zurich)

Faster Machine Learning via Low-Precision Communication & Computation. Dan Alistarh (IST Austria & ETH Zurich), Hantian Zhang (ETH Zurich) Faster Machine Learning via Low-Precision Communication & Computation Dan Alistarh (IST Austria & ETH Zurich), Hantian Zhang (ETH Zurich) 2 How many bits do you need to represent a single number in machine

More information

CS 229 Project Final Report: Reinforcement Learning for Neural Network Architecture Category : Theory & Reinforcement Learning

CS 229 Project Final Report: Reinforcement Learning for Neural Network Architecture Category : Theory & Reinforcement Learning CS 229 Project Final Report: Reinforcement Learning for Neural Network Architecture Category : Theory & Reinforcement Learning Lei Lei Ruoxuan Xiong December 16, 2017 1 Introduction Deep Neural Network

More information

Exploring the Granularity of Sparsity in Convolutional Neural Networks

Exploring the Granularity of Sparsity in Convolutional Neural Networks Exploring the Granularity of Sparsity in Convolutional Neural Networks Anonymous TMCV submission Abstract Sparsity helps reducing the computation complexity of DNNs by skipping the multiplication with

More information

Deep Residual. Variations

Deep Residual. Variations Deep Residual Network and Its Variations Diyu Yang (Originally prepared by Kaiming He from Microsoft Research) Advantages of Depth Degradation Problem Possible Causes? Vanishing/Exploding Gradients. Overfitting

More information