scispace - formally typeset
Search or ask a question
Author

William J. Dally

Other affiliations: California Institute of Technology, Huawei, Nvidia  ...read more
Bio: William J. Dally is an academic researcher from Stanford University. The author has contributed to research in topics: Stream processing & Cache. The author has an hindex of 92, co-authored 451 publications receiving 59303 citations. Previous affiliations of William J. Dally include California Institute of Technology & Huawei.


Papers
More filters
Proceedings Article
15 Feb 2016
TL;DR: Deep Compression as mentioned in this paper proposes a three-stage pipeline: pruning, quantization, and Huffman coding to reduce the storage requirement of neural networks by 35x to 49x without affecting their accuracy.
Abstract: Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. To address this limitation, we introduce "deep compression", a three stage pipeline: pruning, trained quantization and Huffman coding, that work together to reduce the storage requirement of neural networks by 35x to 49x without affecting their accuracy. Our method first prunes the network by learning only the important connections. Next, we quantize the weights to enforce weight sharing, finally, we apply Huffman coding. After the first two steps we retrain the network to fine tune the remaining connections and the quantized centroids. Pruning, reduces the number of connections by 9x to 13x; Quantization then reduces the number of bits that represent each connection from 32 to 5. On the ImageNet dataset, our method reduced the storage required by AlexNet by 35x, from 240MB to 6.9MB, without loss of accuracy. Our method reduced the size of VGG-16 by 49x from 552MB to 11.3MB, again with no loss of accuracy. This allows fitting the model into on-chip SRAM cache rather than off-chip DRAM memory. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. Benchmarked on CPU, GPU and mobile GPU, compressed network has 3x to 4x layerwise speedup and 3x to 7x better energy efficiency.

7,256 citations

Posted Content
TL;DR: This work proposes a small DNN architecture called SqueezeNet, which achieves AlexNet-level accuracy on ImageNet with 50x fewer parameters and is able to compress to less than 0.5MB (510x smaller than AlexNet).
Abstract: Recent research on deep neural networks has focused primarily on improving accuracy. For a given accuracy level, it is typically possible to identify multiple DNN architectures that achieve that accuracy level. With equivalent accuracy, smaller DNN architectures offer at least three advantages: (1) Smaller DNNs require less communication across servers during distributed training. (2) Smaller DNNs require less bandwidth to export a new model from the cloud to an autonomous car. (3) Smaller DNNs are more feasible to deploy on FPGAs and other hardware with limited memory. To provide all of these advantages, we propose a small DNN architecture called SqueezeNet. SqueezeNet achieves AlexNet-level accuracy on ImageNet with 50x fewer parameters. Additionally, with model compression techniques we are able to compress SqueezeNet to less than 0.5MB (510x smaller than AlexNet). The SqueezeNet architecture is available for download here: this https URL

5,904 citations

Proceedings Article
07 Dec 2015
TL;DR: In this paper, the authors proposed a method to reduce the storage and computation required by neural networks by an order of magnitude without affecting their accuracy by learning only the important connections using a three-step method.
Abstract: Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems. Also, conventional networks fix the architecture before training starts; as a result, training cannot improve the architecture. To address these limitations, we describe a method to reduce the storage and computation required by neural networks by an order of magnitude without affecting their accuracy by learning only the important connections. Our method prunes redundant connections using a three-step method. First, we train the network to learn which connections are important. Next, we prune the unimportant connections. Finally, we retrain the network to fine tune the weights of the remaining connections. On the ImageNet dataset, our method reduced the number of parameters of AlexNet by a factor of 9x, from 61 million to 6.7 million, without incurring accuracy loss. Similar experiments with VGG-16 found that the total number of parameters can be reduced by 13x, from 138 million to 10.3 million, again with no loss of accuracy.

3,967 citations

Book
01 Jan 2004
TL;DR: This book offers a detailed and comprehensive presentation of the basic principles of interconnection network design, clearly illustrating them with numerous examples, chapter exercises, and case studies, allowing a designer to see all the steps of the process from abstract design to concrete implementation.
Abstract: One of the greatest challenges faced by designers of digital systems is optimizing the communication and interconnection between system components. Interconnection networks offer an attractive and economical solution to this communication crisis and are fast becoming pervasive in digital systems. Current trends suggest that this communication bottleneck will be even more problematic when designing future generations of machines. Consequently, the anatomy of an interconnection network router and science of interconnection network design will only grow in importance in the coming years. This book offers a detailed and comprehensive presentation of the basic principles of interconnection network design, clearly illustrating them with numerous examples, chapter exercises, and case studies. It incorporates hardware-level descriptions of concepts, allowing a designer to see all the steps of the process from abstract design to concrete implementation. ·Case studies throughout the book draw on extensive author experience in designing interconnection networks over a period of more than twenty years, providing real world examples of what works, and what doesn't. ·Tightly couples concepts with implementation costs to facilitate a deeper understanding of the tradeoffs in the design of a practical network. ·A set of examples and exercises in every chapter help the reader to fully understand all the implications of every design decision. Table of Contents Chapter 1 Introduction to Interconnection Networks 1.1 Three Questions About Interconnection Networks 1.2 Uses of Interconnection Networks 1.3 Network Basics 1.4 History 1.5 Organization of this Book Chapter 2 A Simple Interconnection Network 2.1 Network Specifications and Constraints 2.2 Topology 2.3 Routing 2.4 Flow Control 2.5 Router Design 2.6 Performance Analysis 2.7 Exercises Chapter 3 Topology Basics 3.1 Nomenclature 3.2 Traffic Patterns 3.3 Performance 3.4 Packaging Cost 3.5 Case Study: The SGI Origin 2000 3.6 Bibliographic Notes 3.7 Exercises Chapter 4 Butterfly Networks 4.1 The Structure of Butterfly Networks 4.2 Isomorphic Butterflies 4.3 Performance and Packaging Cost 4.4 Path Diversity and Extra Stages 4.5 Case Study: The BBN Butterfly 4.6 Bibliographic Notes 4.7 Exercises Chapter 5 Torus Networks 5.1 The Structure of Torus Networks 5.2 Performance 5.3 Building Mesh and Torus Networks 5.4 Express Cubes 5.5 Case Study: The MIT J-Machine 5.6 Bibliographic Notes 5.7 Exercises Chapter 6 Non-Blocking Networks 6.1 Non-Blocking vs. Non-Interfering Networks 6.2 Crossbar Networks 6.3 Clos Networks 6.4 Benes Networks 6.5 Sorting Networks 6.6 Case Study: The Velio VC2002 (Zeus) Grooming Switch 6.7 Bibliographic Notes 6.8 Exercises Chapter 7 Slicing and Dicing 7.1 Concentrators and Distributors 7.2 Slicing and Dicing 7.3 Slicing Multistage Networks 7.4 Case Study: Bit Slicing in the Tiny Tera 7.5 Bibliographic Notes 7.6 Exercises Chapter 8 Routing Basics 8.1 A Routing Example 8.2 Taxonomy of Routing Algorithms 8.3 The Routing Relation 8.4 Deterministic Routing 8.5 Case Study: Dimension-Order Routing in the Cray T3D 8.6 Bibliographic Notes 8.7 Exercises Chapter 9 Oblivious Routing 9.1 Valiant's Randomized Routing Algorithm 9.2 Minimal Oblivious Routing 9.3 Load-Balanced Oblivious Routing 9.4 Analysis of Oblivious Routing 9.5 Case Study: Oblivious Routing in the Avici Terabit Switch Router(TSR) 9.6 Bibliographic Notes 9.7 Exercises Chapter 10 Adaptive Routing 10.1 Adaptive Routing Basics 10.2 Minimal Adaptive Routing 10.3 Fully Adaptive Routing 10.4 Load-Balanced Adaptive Routing 10.5 Search-Based Routing 10.6 Case Study: Adaptive Routing in the Thinking Machines CM-5 10.7 Bibliographic Notes 10.8 Exercises Chapter 11 Routing Mechanics 11.1 Table-Based Routing 11.2 Algorithmic Routing 11.3 Case Study: Oblivious Source Routing in the IBM Vulcan Network 11.4 Bibliographic Notes 11.5 Exercises Chapter 12 Flow Control Basics 12.1 Resources and Allocation Units 12.2 Bufferless Flow Control 12.3 Circuit Switching 12.4 Bibliographic Notes 12.5 Exercises Chapter 13 Buffered Flow Control 13.1 Packet-Buffer Flow Control 13.2 Flit-Buffer Flow Control 13.3 Buffer Management and Backpressure 13.4 Flit-Reservation Flow Control 13.5 Bibliographic Notes 13.6 Exercises Chapter 14 Deadlock and Livelock 14.1 Deadlock 14.2 Deadlock Avoidance 14.3 Adaptive Routing 14.4 Deadlock Recovery 14.5 Livelock 14.6 Case Study: Deadlock Avoidance in the Cray T3E 14.7 Bibliographic Notes 14.8 Exercises Chapter 15 Quality of Service 15.1 Service Classes and Service Contracts 15.2 Burstiness and Network Delays 15.3 Implementation of Guaranteed Services 15.4 Implementation of Best-Effort Services 15.5 Separation of Resources 15.6 Case Study: ATM Service Classes 15.7 Case Study: Virtual Networks in the Avici TSR 15.8 Bibliographic Notes 15.9 Exercises Chapter 16 Router Architecture 16.1 Basic Router Architecture 16.2 Stalls 16.3 Closing the Loop with Credits 16.4 Reallocating a Channel 16.5 Speculation and Lookahead 16.6 Flit and Credit Encoding 16.7 Case Study: The Alpha 21364 Router 16.8 Bibliographic Notes 16.9 Exercises Chapter 17 Router Datapath Components 17.1 Input Buffer Organization 17.2 Switches 17.3 Output Organization 17.4 Case Study: The Datapath of the IBM Colony Router 17.5 Bibliographic Notes 17.6 Exercises Chapter 18 Arbitration 18.1 Arbitration Timing 18.2 Fairness 18.3 Fixed Priority Arbiter 18.4 Variable Priority Iterative Arbiters 18.5 Matrix Arbiter 18.6 Queuing Arbiter 18.7 Exercises Chapter 19 Allocation 19.1 Representations 19.2 Exact Algorithms 19.3 Separable Allocators 19.4 Wavefront Allocator 19.5 Incremental vs. Batch Allocation 19.6 Multistage Allocation 19.7 Performance of Allocators 19.8 Case Study: The Tiny Tera Allocator 19.9 Bibliographic Notes 19.10 Exercises Chapter 20 Network Interfaces 20.1 Processor-Network Interface 20.2 Shared-Memory Interface 20.3 Line-Fabric Interface 20.4 Case Study: The MIT M-Machine Network Interface 20.5 Bibliographic Notes 20.6 Exercises Chapter 21 Error Control 411 21.1 Know Thy Enemy: Failure Modes and Fault Models 21.2 The Error Control Process: Detection, Containment, and Recovery 21.3 Link Level Error Control 21.4 Router Error Control 21.5 Network-Level Error Control 21.6 End-to-end Error Control 21.7 Bibliographic Notes 21.8 Exercises Chapter 22 Buses 22.1 Bus Basics 22.2 Bus Arbitration 22.3 High Performance Bus Protocol 22.4 From Buses to Networks 22.5 Case Study: The PCI Bus 22.6 Bibliographic Notes 22.7 Exercises Chapter 23 Performance Analysis 23.1 Measures of Interconnection Network Performance 23.2 Analysis 23.3 Validation 23.4 Case Study: Efficiency and Loss in the BBN Monarch Network 23.5 Bibliographic Notes 23.6 Exercises Chapter 24 Simulation 24.1 Levels of Detail 24.2 Network Workloads 24.3 Simulation Measurements 24.4 Simulator Design 24.5 Bibliographic Notes 24.6 Exercises Chapter 25 Simulation Examples 495 25.1 Routing 25.2 Flow Control Performance 25.3 Fault Tolerance Appendix A Nomenclature Appendix B Glossary Appendix C Network Simulator

3,233 citations

Proceedings ArticleDOI
22 Jun 2001
TL;DR: This paper introduces the concept of on-chip networks, sketches a simple network, and discusses some challenges in the architecture and design of these networks.
Abstract: Using on-chip interconnection networks in place of ad-hoc glo-bal wiring structures the top level wires on a chip and facilitates modular design. With this approach, system modules (processors, memories, peripherals, etc...) communicate by sending packets to one another over the network. The structured network wiring gives well-controlled electrical parameters that eliminate timing iterations and enable the use of high-performance circuits to reduce latency and increase bandwidth. The area overhead required to implement an on-chip network is modest, we estimate 6.6%. This paper introduces the concept of on-chip networks, sketches a simple network, and discusses some challenges in the architecture and design of these networks.

3,209 citations


Cited by
More filters
Posted Content
TL;DR: This work introduces two simple global hyper-parameters that efficiently trade off between latency and accuracy and demonstrates the effectiveness of MobileNets across a wide range of applications and use cases including object detection, finegrain classification, face attributes and large scale geo-localization.
Abstract: We present a class of efficient models called MobileNets for mobile and embedded vision applications. MobileNets are based on a streamlined architecture that uses depth-wise separable convolutions to build light weight deep neural networks. We introduce two simple global hyper-parameters that efficiently trade off between latency and accuracy. These hyper-parameters allow the model builder to choose the right sized model for their application based on the constraints of the problem. We present extensive experiments on resource and accuracy tradeoffs and show strong performance compared to other popular models on ImageNet classification. We then demonstrate the effectiveness of MobileNets across a wide range of applications and use cases including object detection, finegrain classification, face attributes and large scale geo-localization.

14,406 citations

Journal ArticleDOI
01 Apr 1988-Nature
TL;DR: In this paper, a sedimentological core and petrographic characterisation of samples from eleven boreholes from the Lower Carboniferous of Bowland Basin (Northwest England) is presented.
Abstract: Deposits of clastic carbonate-dominated (calciclastic) sedimentary slope systems in the rock record have been identified mostly as linearly-consistent carbonate apron deposits, even though most ancient clastic carbonate slope deposits fit the submarine fan systems better. Calciclastic submarine fans are consequently rarely described and are poorly understood. Subsequently, very little is known especially in mud-dominated calciclastic submarine fan systems. Presented in this study are a sedimentological core and petrographic characterisation of samples from eleven boreholes from the Lower Carboniferous of Bowland Basin (Northwest England) that reveals a >250 m thick calciturbidite complex deposited in a calciclastic submarine fan setting. Seven facies are recognised from core and thin section characterisation and are grouped into three carbonate turbidite sequences. They include: 1) Calciturbidites, comprising mostly of highto low-density, wavy-laminated bioclast-rich facies; 2) low-density densite mudstones which are characterised by planar laminated and unlaminated muddominated facies; and 3) Calcidebrites which are muddy or hyper-concentrated debrisflow deposits occurring as poorly-sorted, chaotic, mud-supported floatstones. These

9,929 citations

Proceedings ArticleDOI
Mark Sandler1, Andrew Howard1, Menglong Zhu1, Andrey Zhmoginov1, Liang-Chieh Chen1 
18 Jun 2018
TL;DR: MobileNetV2 as mentioned in this paper is based on an inverted residual structure where the shortcut connections are between the thin bottleneck layers and intermediate expansion layer uses lightweight depthwise convolutions to filter features as a source of non-linearity.
Abstract: In this paper we describe a new mobile architecture, MobileNetV2, that improves the state of the art performance of mobile models on multiple tasks and benchmarks as well as across a spectrum of different model sizes. We also describe efficient ways of applying these mobile models to object detection in a novel framework we call SSDLite. Additionally, we demonstrate how to build mobile semantic segmentation models through a reduced form of DeepLabv3 which we call Mobile DeepLabv3. is based on an inverted residual structure where the shortcut connections are between the thin bottleneck layers. The intermediate expansion layer uses lightweight depthwise convolutions to filter features as a source of non-linearity. Additionally, we find that it is important to remove non-linearities in the narrow layers in order to maintain representational power. We demonstrate that this improves performance and provide an intuition that led to this design. Finally, our approach allows decoupling of the input/output domains from the expressiveness of the transformation, which provides a convenient framework for further analysis. We measure our performance on ImageNet [1] classification, COCO object detection [2], VOC image segmentation [3]. We evaluate the trade-offs between accuracy, and number of operations measured by multiply-adds (MAdd), as well as actual latency, and the number of parameters.

9,381 citations

Posted Content
Mark Sandler1, Andrew Howard1, Menglong Zhu1, Andrey Zhmoginov1, Liang-Chieh Chen1 
TL;DR: A new mobile architecture, MobileNetV2, is described that improves the state of the art performance of mobile models on multiple tasks and benchmarks as well as across a spectrum of different model sizes and allows decoupling of the input/output domains from the expressiveness of the transformation.
Abstract: In this paper we describe a new mobile architecture, MobileNetV2, that improves the state of the art performance of mobile models on multiple tasks and benchmarks as well as across a spectrum of different model sizes. We also describe efficient ways of applying these mobile models to object detection in a novel framework we call SSDLite. Additionally, we demonstrate how to build mobile semantic segmentation models through a reduced form of DeepLabv3 which we call Mobile DeepLabv3. The MobileNetV2 architecture is based on an inverted residual structure where the input and output of the residual block are thin bottleneck layers opposite to traditional residual models which use expanded representations in the input an MobileNetV2 uses lightweight depthwise convolutions to filter features in the intermediate expansion layer. Additionally, we find that it is important to remove non-linearities in the narrow layers in order to maintain representational power. We demonstrate that this improves performance and provide an intuition that led to this design. Finally, our approach allows decoupling of the input/output domains from the expressiveness of the transformation, which provides a convenient framework for further analysis. We measure our performance on Imagenet classification, COCO object detection, VOC image segmentation. We evaluate the trade-offs between accuracy, and number of operations measured by multiply-adds (MAdd), as well as the number of parameters

8,807 citations

Proceedings Article
15 Feb 2016
TL;DR: Deep Compression as mentioned in this paper proposes a three-stage pipeline: pruning, quantization, and Huffman coding to reduce the storage requirement of neural networks by 35x to 49x without affecting their accuracy.
Abstract: Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. To address this limitation, we introduce "deep compression", a three stage pipeline: pruning, trained quantization and Huffman coding, that work together to reduce the storage requirement of neural networks by 35x to 49x without affecting their accuracy. Our method first prunes the network by learning only the important connections. Next, we quantize the weights to enforce weight sharing, finally, we apply Huffman coding. After the first two steps we retrain the network to fine tune the remaining connections and the quantized centroids. Pruning, reduces the number of connections by 9x to 13x; Quantization then reduces the number of bits that represent each connection from 32 to 5. On the ImageNet dataset, our method reduced the storage required by AlexNet by 35x, from 240MB to 6.9MB, without loss of accuracy. Our method reduced the size of VGG-16 by 49x from 552MB to 11.3MB, again with no loss of accuracy. This allows fitting the model into on-chip SRAM cache rather than off-chip DRAM memory. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. Benchmarked on CPU, GPU and mobile GPU, compressed network has 3x to 4x layerwise speedup and 3x to 7x better energy efficiency.

7,256 citations