Deep In-memory Architectures for Machine Learning

1 Introduction 1.1 The Energy Problem in Machine Learning 1.2 Digital ML Architectures 1.3 In-memory ML Architectures 1.4 Book Organization 2 The Deep In-memory Architecture (DIMA) 2.1 Data-flow of Machine Learning Algorithms 2.2 DIMA Overview 2.3 Inference Architectures: A Shannon-inspired Perspective 2.4 DIMA Design Guidelines and Techniques 2.5 DIMA Models of Energy, Delay, and Accuracy 2.6 Conclusion Appendices 3 DIMA Prototype Integrated Circuits 3.1 The Multi-Functional DIMA IC 3.2 Measured Results 3.3 Random Forest (RF) DIMA IC 3.4 Random Forest IC Prototype 3.5 Measured Results 3.6 Conclusion 4 A Variation-Tolerant DIMA via On-Chip Training 4.1 Background and Rationale 4.2 Architecture and Circuit Implementation 4.3 Experimental Results 4.4 Conclusion 5 Mapping Inference Algorithms to DIMA 5.1 Convolutional Neural Network (CNN) 5.2 Mapping CNN on DIMA (DIMA-CNN) 5.3 Energy, Delay, and Functional Models of DIMA-CNN 5.4 Simulation and Results 5.5 Sparse Distributed Memory (SDM) 5.6 DIMA-based SDM Architecture (DIMA-SDM) 5.7 Energy, Delay, and Functional Models of DIMA-SDM 5.8 Simulation Results 5.9 Conclusions 6 PROMISE: A DIMA-based Accelerator 6.1 Background 6.2 DIMA Instruction Set Architecture 6.3 Compiler 6.4 Validation Methodology 6.5 Evaluation 6.6 Conclusion 7 Future Prospects Index