Spiking Neural Networks

Report on Recent Developments in Spiking Neural Networks

General Direction of the Field

The field of Spiking Neural Networks (SNNs) is witnessing significant advancements, particularly in enhancing their efficiency, robustness, and applicability across various domains. Recent developments focus on improving the convergence and accuracy of SNNs, especially in sequence learning tasks, by leveraging novel conversion techniques from conventional neural networks. Additionally, there is a growing emphasis on formal verification methods to ensure the reliability and safety of SNNs, addressing concerns related to power consumption and adversarial robustness.

Innovations in encoding methods and network architectures are also prominent, with researchers exploring adaptive and hybrid coding approaches to better utilize the spatiotemporal properties of SNNs. These advancements aim to enhance the performance of SNNs in handling complex, real-world data while maintaining energy efficiency. Furthermore, the integration of context-sensitive dendrites and adaptive tokens in SNN-based vision transformers is demonstrating improved learning capabilities and reduced computational demands.

Noteworthy Developments

  • Optimal SNN in Sequence Learning: Achieving high accuracy in sequence tasks through innovative conversion pipelines and s-analog encoding methods.
  • Efficient Formal Verification of SNNs: Introducing temporal encoding to verify adversarial robustness at practical scales, advancing the safe application of SNNs.
  • Adaptive Spiking Neural Networks with Hybrid Coding: Enhancing the utilization of temporal encoding and reducing training time steps, significantly improving classification performance.
  • Robust Iterative Value Conversion in DRL: Proposing a method to reduce and robustify conversion errors in deep reinforcement learning for neurochip-driven edge robots, leading to substantial power savings and increased calculation speed.

These developments highlight the ongoing efforts to refine SNNs, making them more accurate, efficient, and reliable for a wide range of applications.

Sources

Obtaining Optimal Spiking Neural Network in Sequence Learning via CRNN-SNN Conversion

Towards Efficient Formal Verification of Spiking Neural Network

Achieving the Tightest Relaxation of Sigmoids for Formal Verification

An Overlooked Role of Context-Sensitive Dendrites

Adaptive Spiking Neural Networks with Hybrid Coding

AT-SNN: Adaptive Tokens for Vision Transformer on Spiking Neural Network

Robust Iterative Value Conversion: Deep Reinforcement Learning for Neurochip-driven Edge Robots

Verification of Geometric Robustness of Neural Networks via Piecewise Linear Approximation and Lipschitz Optimisation

Growing Deep Neural Network Considering with Similarity between Neurons

Evolvable Psychology Informed Neural Network for Memory Behavior Modeling

Towards Automatic Linearization via SMT Solving

Built with on top of