Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks

Lihao Wang, Zhaofei Yu

TL;DR

This paper tackles the difficulty of capturing long-term temporal dependencies and rich spatial interactions in spiking neural networks by introducing STC-LIF, an autapse-inspired self-connection that adds two dynamic pathways to regulate input current and historical membrane information. The STC-LIF framework uses modulators built from previous-time-step activity and group convolutions to enable local spatio-temporal interactions, improving memory retention and gradient flow. Across Moving MNIST, TaxiBJ, and KTH benchmarks, STC-LIF and its enhanced variants outperform vanilla LIF and existing adaptive models, demonstrating state-of-the-art spatio-temporal predictive performance and compatibility with other SNN architectures. The approach offers a general, scalable way to enrich SNN dynamics for dynamic tasks and neuromorphic data processing, with potential applicability to broader adaptive neuron models.

Abstract

Spiking Neural Networks (SNNs) emulate the integrated-fire-leak mechanism found in biological neurons, offering a compelling combination of biological realism and energy efficiency. In recent years, they have gained considerable research interest. However, existing SNNs predominantly rely on the Leaky Integrate-and-Fire (LIF) model and are primarily suited for simple, static tasks. They lack the ability to effectively model long-term temporal dependencies and facilitate spatial information interaction, which is crucial for tackling complex, dynamic spatio-temporal prediction tasks. To tackle these challenges, this paper draws inspiration from the concept of autaptic synapses in biology and proposes a novel Spatio-Temporal Circuit (STC) model. The STC model integrates two learnable adaptive pathways, enhancing the spiking neurons' temporal memory and spatial coordination. We conduct a theoretical analysis of the dynamic parameters in the STC model, highlighting their contribution in establishing long-term memory and mitigating the issue of gradient vanishing. Through extensive experiments on multiple spatio-temporal prediction datasets, we demonstrate that our model outperforms other adaptive models. Furthermore, our model is compatible with existing spiking neuron models, thereby augmenting their dynamic representations. In essence, our work enriches the specificity and topological complexity of SNNs.

Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks

TL;DR

This paper tackles the difficulty of capturing long-term temporal dependencies and rich spatial interactions in spiking neural networks by introducing STC-LIF, an autapse-inspired self-connection that adds two dynamic pathways to regulate input current and historical membrane information. The STC-LIF framework uses modulators built from previous-time-step activity and group convolutions to enable local spatio-temporal interactions, improving memory retention and gradient flow. Across Moving MNIST, TaxiBJ, and KTH benchmarks, STC-LIF and its enhanced variants outperform vanilla LIF and existing adaptive models, demonstrating state-of-the-art spatio-temporal predictive performance and compatibility with other SNN architectures. The approach offers a general, scalable way to enrich SNN dynamics for dynamic tasks and neuromorphic data processing, with potential applicability to broader adaptive neuron models.

Abstract

Spiking Neural Networks (SNNs) emulate the integrated-fire-leak mechanism found in biological neurons, offering a compelling combination of biological realism and energy efficiency. In recent years, they have gained considerable research interest. However, existing SNNs predominantly rely on the Leaky Integrate-and-Fire (LIF) model and are primarily suited for simple, static tasks. They lack the ability to effectively model long-term temporal dependencies and facilitate spatial information interaction, which is crucial for tackling complex, dynamic spatio-temporal prediction tasks. To tackle these challenges, this paper draws inspiration from the concept of autaptic synapses in biology and proposes a novel Spatio-Temporal Circuit (STC) model. The STC model integrates two learnable adaptive pathways, enhancing the spiking neurons' temporal memory and spatial coordination. We conduct a theoretical analysis of the dynamic parameters in the STC model, highlighting their contribution in establishing long-term memory and mitigating the issue of gradient vanishing. Through extensive experiments on multiple spatio-temporal prediction datasets, we demonstrate that our model outperforms other adaptive models. Furthermore, our model is compatible with existing spiking neuron models, thereby augmenting their dynamic representations. In essence, our work enriches the specificity and topological complexity of SNNs.
Paper Structure (24 sections, 12 equations, 6 figures, 11 tables)

This paper contains 24 sections, 12 equations, 6 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Learning of Spiking Neural Networks
  4. Adaptive Spiking Neural Models
  5. Preliminaries
  6. Vanilla LIF Model
  7. Autapses
  8. Spatio-temporal Predictive Learning
  9. Methodology
  10. Limitations of the Vanilla LIF Model in Spatio-temporal Prediction Tasks
  11. Spatio-Temporal Circuit Leaky Integrate-and-Fire (STC-LIF) Model
  12. Dynamic Analysis of STC-LIF Model
  13. Experiments
  14. Evaluation on the Moving MNIST Dataset
  15. Evaluation on the TaxiBJ Dataset
  16. ...and 9 more sections

Figures (6)

  • Figure 1: (a) The structure of Autapses in neuroscience, which includes axon-soma and axon-dendrite circuits. (b) The structure of the vanilla LIF model. (c) The structure of the STC-LIF model. (d) The structure of dynamic spatio-temporal circuit module $\boldsymbol \sigma$ in Figure 1(c).
  • Figure 2: Comparison of parameter changes of different models.
  • Figure 3: Qualitative visualization of the prediction results of vanilla LIF and STC-LIF models on the Moving MNIST dataset.
  • Figure 4: Comparison of vanilla LIF (a) and STC-LIF (b) when inputting sequential sequences or random sequences from the DVS128 Gesture dataset. The experiments were repeated three times, and the solid curve represents the average value.
  • Figure 5: The unfolded computation graph of STC-LIF model.
  • ...and 1 more figures