Biologically Inspired Spatial–Temporal Perceiving Strategies for Spiking Neural Network

Yu Zheng, Jingfeng Xue, Jing Liu, Yanjun ZhangBiomimetics 2025 · DOI

TL;DR

This paper proposes a novel approach to enhance the spatial-temporal perception capabilities of Spiking Neural Networks (SNNs) by introducing a neuron group-based structural learning mechanism and a time-slicing interpretation method. The goal is to make AI more interpretable and understandable, moving beyond the "black box" nature of traditional deep learning models, particularly for future autonomous systems.

What Problem Does This Paper Address?

The increasing complexity of real-world environments demands autonomous systems that can perceive, decide, and control effectively, even with incomplete information. Current AI, largely based on Deep Neural Networks (DNNs), struggles with interpretability – they are often "black boxes" where the decision-making process is opaque. This lack of transparency hinders human-AI interaction and trust. Spiking Neural Networks (SNNs), inspired by the biological brain, offer a potential solution due to their inherent temporal processing capabilities. However, effectively leveraging SNNs for robust and interpretable environmental perception, especially capturing both spatial and temporal information, remains a challenge.

Methods

The authors introduce three core contributions:

  1. ROC-based Encoding: A Rank Order Coding (ROC) method is employed to encode pixel values into spike arrival times. This prioritizes more important information by transmitting it earlier, aiming to better capture spatial and temporal dynamics.
  2. Neuron Group Structural Learning: This mechanism focuses on learning structural connections within the SNN. It involves:
  3. Neuron Pair Formation: Based on Hebbian principles, neurons that fire together within a certain time and distance threshold form pairs.
  4. Neuron Pair Refinement: Synapse weights are trained and refined to select pairs that are discriminative for specific digits.
  5. Neuron Group Formation: These refined pairs are then used to form neuron groups, which are connected to form the network's memory layers. This structural learning aims to enhance the SNN's ability to differentiate and store spatial-temporal information.
  6. Time-Slicing Mechanism: To interpret the SNN's responses, the temporal window of spike activity is divided into multiple slices. Each slice generates a "spiking map," and a collection of these maps over time provides a richer representation of the SNN's internal state and decision-making process, enhancing interpretability.

Datasets

The primary dataset used for training and evaluation is the MNIST dataset, a standard benchmark for handwritten digit recognition. The paper also mentions experiments on MNIST-C, a corrupted version of MNIST designed to test robustness against out-of-distribution conditions.

Key Insights and Ideas

  • SNNs as Interpretable AI: The paper strongly advocates for SNNs as a pathway to more interpretable AI, contrasting them with the opaque nature of DNNs.
  • Spatial-Temporal Fusion is Key: Effectively capturing and processing both spatial and temporal information simultaneously is crucial for robust environmental perception, especially for autonomous systems.
  • Structure Matters for SNNs: Unlike fixed DNN architectures, the proposed neuron group learning allows the SNN to dynamically adapt its structure, enhancing its ability to learn complex spatio-temporal patterns.
  • Time Slicing for Transparency: Breaking down the temporal response of an SNN into discrete time slices provides a more granular view of its internal processing, making it easier to understand how it arrives at a decision.
  • Neuron Pair Density is Critical: The number of neuron pairs in the network significantly impacts accuracy, suggesting a sweet spot for effective information representation.

Key Results

  • The proposed neuron group-based structural learning and time-slicing mechanism enhances the environment perception ability of the SNN.
  • The method demonstrates a certain degree of robustness, particularly against noise (accuracy decreases only slightly for noise ratios under 20%).
  • Experiments on MNIST showed that accuracy generally improves with more time slices, especially in the 3-layer network.
  • The number of neuron pairs also plays a crucial role, with accuracy differences between 2-layer and 3-layer networks narrowing as the number of neuron pairs increases.
  • The use of preset currents (sub-excitation) was found to significantly improve accuracy in the 3-layer network, aiding signal propagation.
  • On MNIST-C, the method achieved a reasonably high result compared to other state-of-the-art methods, indicating good performance on corrupted data.

Limitations and Future Work

  • Scalability to Deeper Networks: While the paper presents 2-layer and 3-layer networks, the authors suggest that the neuron group mechanism can be extended to deeper networks, potentially introducing "chain mechanisms and bypass mechanisms" for faster perception.
  • Optimizing Time Slicing: The paper focuses on exploring time slicing for interpretability. Future work could focus on optimizing the selection of slicing time points to maximize information entropy.
  • Neuron Pair Density Optimization: The paper highlights the importance of neuron pair density but doesn't provide a definitive optimal range. Further research could explore more systematic ways to determine optimal neuron pair configurations.
  • Interpretability Metrics: While the time-slicing method aims for interpretability, more quantitative metrics for evaluating interpretability could be explored.

Summary updated

Abstract

A future unmanned system needs the ability to perceive, decide and control in an open dynamic environment. In order to fulfill this requirement, it needs to construct a method with a universal environmental perception ability. Moreover, this perceptual process needs to be interpretable and understandable, so that future interactions between unmanned systems and humans can be unimpeded. However, current mainstream DNN (deep learning neural network)-based AI (artificial intelligence) is a 'black box'. We cannot interpret or understand how the decision is made by these AIs. An SNN (spiking neural network), which is more similar to a biological brain than a DNN, has the potential to implement interpretable or understandable AI. In this work, we propose a neuron group-based structural learning method for an SNN to better capture the spatial and temporal information from the external environment, and propose a time-slicing scheme to better interpret the spatial and temporal information of responses generated by an SNN. Results show that our method indeed helps to enhance the environment perception ability of the SNN, and possesses a certain degree of robustness, enhancing the potential to build an interpretable or understandable AI in the future.
BibTeX
@inproceedings{zheng2025biologically,
  author = {Yu Zheng and Jingfeng Xue and Jing Liu and Yanjun Zhang},
  title = {{Biologically Inspired Spatial–Temporal Perceiving Strategies for Spiking Neural Network}},
  booktitle = {Biomimetics},
  year = {2025},
  doi = {10.3390/biomimetics10010048},
}
Figure