Advances in Spiking Neural Networks

The field of Spiking Neural Networks (SNNs) is rapidly advancing, with a focus on improving energy efficiency, robustness, and accuracy. Recent research has explored innovative approaches to optimize SNN computation, including pattern-based hierarchical sparsity and dominant eigencomponent projection. These methods have shown significant improvements in energy efficiency and robustness compared to traditional SNN accelerators. Another area of research has focused on developing novel hardware architectures and encoding methods to efficiently process SNNs, achieving high accuracy and low inference times. Furthermore, researchers have investigated the application of SNNs in space applications, demonstrating their potential for energy-efficient scene classification.

Noteworthy papers in this area include:

  • Phi, which introduces a novel pattern-based hierarchical sparsity framework to optimize SNN computation, achieving a 3.45x speedup and a 4.93x improvement in energy efficiency.
  • Towards Robust Spiking Neural Networks, which develops a hyperparameter-free method called Dominant Eigencomponent Projection (DEP) to mitigate the vulnerability of SNNs to heterogeneous data poisoning, significantly enhancing overall robustness.
  • Energy-Efficient Deep Reinforcement Learning with Spiking Transformers, which combines the energy efficiency of SNNs with the powerful decision-making capabilities of reinforcement learning, demonstrating significantly improved policy performance and energy efficiency.
  • Adversarially Robust Spiking Neural Networks with Sparse Connectivity, which introduces a neural network conversion algorithm designed to produce sparse and adversarially robust SNNs, achieving state-of-the-art performance with enhanced energy and memory efficiency.

Sources

Phi: Leveraging Pattern-based Hierarchical Sparsity for High-Efficiency Spiking Neural Networks

Towards Robust Spiking Neural Networks:Mitigating Heterogeneous Training Vulnerability via Dominant Eigencomponent Projection

Lightweight LIF-only SNN accelerator using differential time encoding

Energy efficiency analysis of Spiking Neural Networks for space applications

Spiking Neural Networks with Random Network Architecture

Frozen Backpropagation: Relaxing Weight Symmetry in Temporally-Coded Deep Spiking Neural Networks

Energy-Efficient Deep Reinforcement Learning with Spiking Transformers

Adaptive Pruning of Deep Neural Networks for Resource-Aware Embedded Intrusion Detection on the Edge

Adversarially Robust Spiking Neural Networks with Sparse Connectivity

Built with on top of