Cover Image

Towards neuromorphic machine intelligence : spike -based representation, learning, and applications.

Towards Neuromorphic Machine Intelligence explores the field of spiking neural networks (SNNs), the third generation of artificial neural networks that aim to mimic the dynamics of biological neurons more accurately. The book covers fundamental theories, specialized neuron models, and learning algor...

Full description

Bibliographic Details
Main Author: QU, HONG
Corporate Author: ScienceDirect (Online service)
Format: eBook
Language:English
Published: [S.l.] : Academic Press, 2024.
Subjects:
Neural networks (Computer science)
Artificial intelligence.
Réseaux neuronaux (Informatique)
Intelligence artificielle.
artificial intelligence.
Electronic books.
Online Access:Connect to the full text of this electronic book
Table of Contents:
  • Front Cover
  • Towards Neuromorphic Machine Intelligence
  • Copyright
  • Contents
  • Biographies of authors
  • Hong Qu
  • Xiaoling Luo
  • Zhang Yi
  • Foreword
  • Acknowledgments
  • 1 Introduction
  • 1.1 Computational theory of the brain
  • 1.2 What is neuromorphic computing
  • 1.3 Artificial intelligence with neuromorphic computing
  • References
  • 2 Fundamentals of spiking neural networks
  • 2.1 Biological background
  • 2.1.1 Basic structure and information transmission mechanism of biological neurons
  • 2.1.2 The generation of action potentials: spikes
  • 2.2 Neuronal diversity and spiking neuron models
  • 2.2.1 Hodgkin and Huxley model
  • 2.2.2 Leaky integrate-and-fire model
  • 2.2.3 Spike response model
  • 2.3 Conclusion
  • References
  • 3 Specialized spiking neuron model
  • 3.1 Resonate spiking neuron
  • 3.2 Rectified linear PSP-based spiking neuron
  • 3.3 Pyramidal structure spiking neuron
  • 3.4 Signed neuron with memory
  • 3.5 Conclusion
  • References
  • 4 Learning algorithms for shallow spiking neural networks
  • 4.1 Preface
  • 4.2 Accuracy measure and neuron model
  • 4.2.1 Spike metric
  • 4.2.2 Neuron model
  • 4.3 First error-based supervised learning algorithm
  • 4.3.1 First error learning rule
  • 4.3.2 Extended first error learning rule
  • 4.3.2.1 Synaptic delay learning
  • 4.3.2.2 Multi-layer network learning
  • 4.3.3 Algorithm analysis
  • 4.3.3.1 Replace tko with tkd when calculating the gradient
  • 4.3.3.2 Control the influence of spike before te on weight update
  • 4.3.3.3 Initialize parameters appropriately
  • 4.3.3.4 Avoid gradient explosion
  • 4.3.4 Applications and results
  • 4.3.4.1 Performance evaluation of FE-learn
  • 4.3.4.2 Performance evaluation of extended FE-learn
  • 4.3.4.3 Classification of UCI datasets
  • 4.3.4.4 Classification of multimodal data
  • 4.4 Recursive least squares-based learning algorithm.
  • 4.4.1 RLSBLR learning rule
  • 4.4.1.1 The weight learning rule
  • 4.4.1.2 The delay learning rule
  • 4.4.2 Comparison with the existing learning methods
  • 4.4.3 Applications and results
  • 4.4.3.1 Performance evaluation of RLSBLR
  • 4.4.3.2 Classification applications and results
  • 4.5 Membrane potential driven aggregate-label learning
  • 4.5.1 MPD-Al learning algorithm
  • 4.5.1.1 Firing less spikes than desired (No&lt
  • Nd)
  • 4.5.1.2 Firing more spikes than desired (No&gt
  • Nd)
  • 4.5.2 Comparison with other aggregate-label learning algorithms
  • 4.5.3 Applications and results
  • 4.5.3.1 Learning to fire a desired number of spikes
  • 4.5.3.2 Learning predictive clues
  • 4.5.3.3 Application to speech recognition
  • 4.6 Efficient threshold driven aggregate-label learning algorithm
  • 4.6.1 ETDP learning algorithm for single spiking neurons
  • 4.6.2 Extended ETDP learning algorithm for multi-layer SNNs
  • 4.6.3 Applications and results
  • 4.6.3.1 Learning to fire a desired number of spikes
  • 4.6.3.2 Learning multimodal sensory clues
  • 4.6.3.3 Classification tasks
  • 4.7 Conclusions
  • References
  • 5 Learning algorithms for deep spiking neural networks
  • 5.1 Preface
  • 5.1.1 Non-differentiable spike function
  • 5.1.2 Gradient explosion
  • 5.1.3 Dead neuron
  • 5.2 Learning algorithm with neural oscillation and phase information
  • 5.2.1 Solutions to three problems in learning of deep SNNs
  • 5.2.2 Algorithm description
  • 5.2.3 Applications and results
  • 5.2.3.1 XOR problem
  • 5.2.3.2 Vision task
  • 5.3 Learning algorithms with rectified linear postsynaptic potential
  • 5.3.1 Solutions to three problems in learning of deep SNNs
  • 5.3.1.1 Non-differentiable spike function
  • 5.3.1.2 Gradient explosion
  • 5.3.1.3 Dead neuron
  • 5.3.2 Algorithm description
  • 5.3.3 Hardware simulation methodology
  • 5.3.3.1 Hardware architecture.
  • 5.3.3.2 Mapping
  • 5.3.3.3 Hardware simulation methodology
  • 5.3.4 Applications and results
  • 5.3.4.1 Temporal coding
  • 5.3.4.2 MNIST dataset
  • 5.3.4.3 Fashion-MNIST dataset
  • 5.3.4.4 Caltech face/motorbike dataset
  • 5.3.4.5 Hardware simulation results
  • 5.4 Conclusion
  • References
  • 6 Neural column-inspired spiking neural networks for episodic memory
  • 6.1 Preface
  • 6.2 Minicolumn-based model for episodic memory
  • 6.2.1 Model architecture
  • 6.2.1.1 The neuron structure
  • 6.2.1.2 The minicolumn organization
  • 6.2.1.3 The neural networks organization
  • 6.2.2 How does the BSTM simulate episodic memory?
  • 6.2.2.1 The encoding process
  • 6.2.2.2 The storing process
  • 6.2.2.3 The retrieval process
  • 6.2.3 Applications and results
  • 6.2.3.1 The impact of corrupted training sequences on retrieval performance
  • 6.2.3.2 The effect of network parameters on retrieval performance
  • 6.2.3.3 Test memory capacity
  • 6.2.3.4 Comparison with other algorithms
  • 6.3 Columnar-structured model for temporal-sequential learning
  • 6.3.1 Network design
  • 6.3.1.1 Spatiotemporal subnetwork
  • 6.3.1.2 Goal cluster
  • 6.3.1.3 Spiking neuron model inspired by pyramidal cell
  • 6.3.2 Learning and retrieval
  • 6.3.2.1 Choice of winner
  • 6.3.2.2 Establishment of connections
  • 6.3.2.3 Learning process of a sequence
  • 6.3.2.4 Retrieval
  • 6.3.3 Applications and results
  • 6.3.3.1 Goal-based retrieval
  • 6.3.3.2 Context-based retrieval
  • 6.3.3.3 Parameter influence
  • 6.4 Conclusion
  • References
  • 7 An ANN-SNN algorithm suitable for ultra energy efficient image classification
  • 7.1 Preface
  • 7.2 Convert pre-trained ANNs to SNNs
  • 7.2.1 Forward propagation of ANN
  • 7.2.2 Forward propagation of SNN
  • 7.2.3 Converting ANNs to SNNs
  • 7.3 ANN-SNN conversion with signed neuron with memory
  • 7.3.1 Error analysis.
  • 7.3.2 The solution: signed neuron with memory
  • 7.4 Neuron-wise normalization
  • 7.4.1 Problem of long inference latency
  • 7.4.2 The solution: neuron-wise normalization
  • 7.5 Hardware energy consumption analysis
  • 7.6 Applications and results
  • 7.6.1 Details
  • 7.6.2 Ablation study
  • 7.6.3 Comparison with related work
  • 7.7 Conclusions
  • References
  • 8 Spiking deep belief networks for fault diagnosis
  • 8.1 Preface
  • 8.2 Building blocks of spiking deep belief networks
  • 8.2.1 Restricted Boltzmann machine
  • 8.2.2 Siegert neuron model
  • 8.2.3 Neuron coding
  • 8.3 Event-driven spike deep belief network
  • 8.3.1 Network design
  • 8.3.2 Reward-STDP learning rule
  • 8.3.2.1 Event-driven strategy
  • 8.3.2.2 Reward-STDP
  • 8.4 Applications and results
  • 8.4.1 Classification of UCI datasets
  • 8.4.1.1 Data description
  • 8.4.1.2 Data preprocessing
  • 8.4.1.3 Implementation details
  • 8.4.1.4 Evaluation metric
  • 8.4.1.5 Accuracy
  • 8.4.1.6 Ablation experiments
  • 8.4.1.7 Evaluation under different noisy environments
  • 8.4.2 Classification of CWRU datasets
  • 8.4.2.1 Data description
  • 8.4.2.2 Implementation details
  • 8.4.2.3 Visualization of layers
  • 8.4.2.4 Results on the CWRU dataset
  • 8.5 Conclusions
  • References
  • 9 Conclusions
  • 9.1 Summary
  • 9.2 Future outlook
  • Glossary
  • Index
  • Back Cover.