Spiking Neural Networks

Spiking Neural Networks (SNNs) are a class of neural computation models that operate using discrete spike events (“all-or-nothing” action potentials) rather than the continuous-valued activations of conventional Artificial Neural Networks (ANNs). This makes them more biologically plausible, energy-efficient, and naturally suited to temporal/spike-based processing.

Key Properties

• Event-driven computation: Neurons only communicate when their membrane potential crosses a threshold, producing a spike — no computation occurs in the absence of spikes, leading to extreme energy efficiency
• Temporal dynamics: Information is encoded not just in spike rate but also in spike timing, enabling natural processing of time-series data
• Neuromorphic hardware compatibility: SNN architectures map directly onto event-driven neuromorphic chips (e.g., Intel Loihi, IBM TrueNorth), achieving orders-of-magnitude power savings over conventional GPU/CPU implementations

Neuron Models

The Leaky Integrate-and-Fire (LIF) model as the primary focus. LIF neurons integrate incoming spikes over time, allow the membrane potential to decay (leak), and fire an output spike when threshold is reached, after which the potential resets. Luu et al. (2026, IEEE Access) provide a comprehensive taxonomy of spiking neuron models alongside synaptic modelling approaches, giving a useful reference for situating more specialised models such as the covariant receptive field neurons of Pedersen et al.

Training Strategies

Three main approaches are evaluated:

1. Surrogate Gradient Descent — Approximates the non-differentiable spike function with a smooth surrogate during backpropagation. Achieves accuracy within 1-2% of ANNs, with faster convergence (~20 epochs) and latency as low as 10 ms.
2. ANN-to-SNN Conversion — Trains a conventional ANN, then converts activations to spike rates. Competitive performance but requires higher spike counts and longer simulation windows.
3. Spike-Timing Dependent Plasticity (STDP) — Biologically plausible, unsupervised Hebbian learning rule that adjusts synaptic weights based on the relative timing of pre- and post-synaptic spikes. Slowest to converge but lowest energy consumption (~5 mJ per inference), optimal for low-power edge applications.

Frameworks and Tools

The practical development of SNNs has been accelerated by dedicated software frameworks. Chief among these is SpikingJelly (Fang et al., 2023, Science Advances), a full-stack open-source platform for spike-based deep learning that provides:

• Neuromorphic dataset preprocessing — built-in handling of event-based vision datasets (e.g., DVS-CIFAR10, N-MNIST)
• Automatic differentiation for SNNs — surrogate gradient support integrated with PyTorch’s autograd engine, enabling standard backpropagation through spiking neurons
• 11× training acceleration compared to existing implementations, via multilevel inheritance and semiautomatic code generation for custom models
• Neuromorphic chip deployment — direct export to hardware platforms (e.g., Intel Loihi), bridging the gap between simulation and physical neuromorphic systems

SpikingJelly lowered the barrier to entry for SNN research by providing a PyTorch-native API that researchers can use without writing custom CUDA kernels or hand-rolling surrogate gradient functions. Its widespread adoption has enabled many of the advances reviewed in the sections below.

Applications

SNNs are particularly well-suited for:

• Neuromorphic vision — event-based cameras (DVS) naturally produce spike-encoded output; SNNs process these with minimal latency
• Robotics — real-time sensorimotor control on low-power embedded hardware
• Edge AI — inference on battery-powered devices where energy is the binding constraint

Challenges

Despite their promise, SNNs face obstacles:

• Hardware standardisation — no dominant neuromorphic platform has emerged, fragmenting the ecosystem
• Scalable training — surrogate gradient methods work but don’t fully exploit the temporal advantages of spiking dynamics
• Benchmarking — lack of standardised metrics and datasets that fairly compare SNNs to ANNs on temporal tasks

Relevance to Vision Science

SNNs are particularly relevant to computational vision science: they model the event-driven, asynchronous processing of the early visual system (retina → LGN → V1), where spike timing carries information about stimulus onset, motion, and contrast. Event-based cameras (e.g., DAVIS, Prophesee) generate spike-encoded visual streams that SNNs process natively, bridging the gap between eye-tracking hardware and neural computation models.

Covariant Spatio-Temporal Receptive Fields

Pedersen, Conradt, and Lindeberg (2025, Nature Communications) introduced covariant spatio-temporal receptive fields for SNNs — a principled computational model combining affine Gaussian kernels (spatial) with LIF dynamics (temporal) that is provably covariant to spatial affine and temporal scaling transformations. Using these receptive fields as priors for SNN training on event-based vision tasks improves performance beyond standard ANNs, directly addressing the training difficulties identified in the Aribe (2025) review. This work bridges scale-space theory, visual neuroscience, and neuromorphic computing.

V1 Temporal Prediction with Recurrent SNNs

Taylor, Zenke, King, and Harper (2024, bioRxiv) developed a recurrently connected spiking network of excitatory and inhibitory units trained for efficient temporal prediction of natural movie clips — rather than the more common classification or reconstruction objective. The model spontaneously exhibited:

• Simple and complex cell-like tuning — matching the known functional diversity of V1 neurons
• V1-like spike statistics — sparse and irregular firing patterns characteristic of cortical recordings
• E/I differences — the network captured key distinctions between excitatory and inhibitory V1 neurons in their spike statistics, tuning preferences, connectivity, and temporal dynamics, which no single computational principle had previously explained

The fact that these properties emerged from a single objective — efficient temporal prediction of the sensory future — suggests that V1’s excitatory-inhibitory circuitry may be fundamentally organised around predictive coding of visual input over time. This bridges SNN research with the broader predictive coding framework in computational neuroscience.

Topographic Deep SNNs for Visual Cortex Modeling

Zhou, Fang, Wang, and Xu (2025, arXiv) proposed Topographic Deep Spiking Neural Networks (TDSNNs) with a novel Spatio-Temporal Constraints (STC) loss that replicates the hierarchical spatial functional organisation of primate visual cortex — from low-level sensory areas to high-level abstract representations. Key findings:

• Topography without performance cost: Unlike topographic ANNs (e.g., TopoNet, which loses ~3% ImageNet top-1 accuracy), TDSNNs exhibit zero drop in ImageNet top-1 accuracy, while surpassing topographic ANNs in brain-likeness metrics
• Temporal dynamics enable efficient topography: The spike-based temporal processing in SNNs is shown to facilitate stable topographic organisation, contributing to model robustness — a mechanism not available to conventional ANNs
• Hierarchical cortical mimicry: The STC loss successfully generates representative topographic features across simulated visual cortical areas, from V1-like low-level feature maps to higher-level abstract representations

This work demonstrates that SNNs offer a unique advantage over ANNs for neurobiologically constrained vision models: spike-based temporal dynamics allow topographic organisation to emerge without the performance penalty that plagues topographic ANNs. TDSNNs thus provide a computational framework for interpreting topographic organisation in primate visual cortex while maintaining competitive task performance.

BioMotion-SNN: MT-MST Pathways for Visual Motion Perception

Liu, Mei, Feng, Zhang, Qu, and Zhang (2025, Neural Networks) proposed BioMotion-SNN, an SNN framework inspired by the MT-MST (middle temporal–medial superior temporal) visual motion pathways to model the neural processing underlying biological motion perception. Key innovations:

• Event-driven stimulus processing: Unlike conventional models reliant on static datasets and predefined labels, BioMotion-SNN processes event-driven stimuli in a manner closely resembling real sensory inputs
• Contrastive self-supervised learning: A motion-perception contrastive loss function enhances feature representation without requiring manually labelled motion categories
• L1-norm synaptic pruning: Mimics sparse biological connectivity by reducing redundant connections, consistent with known cortical wiring efficiency
• Electrophysiologically grounded: Uses real electrophysiological data augmented through controlled transformations, reducing the need for extensive biological data collection
• Performance: Achieves 93.00% classification accuracy on biological motion tasks

The framework establishes a paradigm for integrating computational SNN modelling with empirical neuroscience by directly incorporating electrophysiological recordings into the training pipeline — bridging the gap between theoretically principled receptive field models and data-driven approaches. Data and code are openly available at github.com/BrainCogLab/biomotion_snn.

See Also

• analog-cognition-consciousness — Hybrid synapse/wave theory: brain waves as an analog computing substrate complementing spike-based models
• ephaptic-coupling-and-brain-waves — Direct (non-synaptic) electric-field interactions that modulate spiking

References

• Aribe, S. G. (2025). Spiking Neural Networks: The Future of Brain-Inspired Computing. arXiv:2510.27379. ¹
• Pedersen, J. E., Conradt, J., & Lindeberg, T. (2025). Covariant spatio-temporal receptive fields for spiking neural networks. Nature Communications, 16, 8231. ²
• Taylor, L., Zenke, F., King, A. J., & Harper, N. (2024). Temporal prediction captures key differences between spiking excitatory and inhibitory V1 neurons. bioRxiv. https://doi.org/10.1101/2024.05.12.593763 ³
• Zhou, D., Fang, Y., Wang, Z., & Xu, R. (2025). TDSNNs: Competitive Topographic Deep Spiking Neural Networks for Visual Cortex Modeling. arXiv:2508.04270. ⁴
• Liu, Y., Mei, J., Feng, T., Zhang, T., Qu, H., & Zhang, Y. (2025). BioMotion-SNN: Spiking neural network modeling for visual motion processing. Neural Networks, 195, 108239. ⁵
• Fang, W., Chen, Y., Ding, J., Yu, Z., Masquelier, T., Chen, D., Huang, L., Zhou, H., Li, G., & Tian, Y. (2023). SpikingJelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9, eadi1480. ⁶
• Luu, N. T., Luu, D. T., Nam, P. N., & Thang, T. C. (2026). A Survey on Spiking Neural Network Foundation and Recent Progress. IEEE Access, 14. https://doi.org/10.1109/ACCESS.2026.3685666 ⁷

Footnotes

1. raw/papers/aribe-2025-spiking-neural-networks.md ↩

2. raw/papers/pedersen-2025-covariant-receptive-fields-snn.md ↩

3. raw/papers/taylor-2024-temporal-prediction-v1-snn.md ↩

4. raw/papers/zhou-2025-tdsnns-topographic-snn-visual-cortex.md ↩

5. raw/papers/liu-2025-biomotion-snn.md ↩

6. raw/papers/fang-2023-spikingjelly.md ↩

7. raw/papers/luu-2026-survey-snn-foundations.md ↩