Research map / Architectures and efficiency / Spiking networks
Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STC-LIF, a spiking neuron with learnable autaptic-inspired spatio-temporal self-connection circuits that modulate input current and membrane potential from the previous spike output, improving spatio-temporal prediction over LIF and adaptive spiking neuron variants.
Problem
Vanilla LIF spiking neurons, and fixed-parameter adaptive variants, struggle to model long-term temporal dependencies, spatial-temporal interaction, and adapt to data for complex spatio-temporal prediction tasks.
Approach
Inspired by autaptic synapses, the neuron's own previous-step spike output (via learnable group convolutions for axon-soma and axon-dendrite pathways) produces dynamic modulation factors that scale historical membrane potential and input current, making neuron dynamics input-adaptive; the same circuit is also added to PLIF and LM-H.
Key results
- On Moving MNIST, STC-LIF reaches MSE 47.0 and SSIM 0.863 versus 102.8 and 0.640 for vanilla LIF and 79.7 for the best prior baseline GLIF.
- On KTH (10→20), STC-LIF achieves SSIM 0.822 and PSNR 23.14 versus 0.805 and 22.41 for GLIF.
- Applying the circuit to PLIF and LM-H improves Moving MNIST MSE from 99.2 to 48.4 (PLIF) and from 93.2 to 45.9 (LM-H).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add spatio-temporal circuit to LIF | LIF · Moving MNIST | MSE: 102.8 -> 47.0 | Helped |
| add spatio-temporal circuit to LIF | LIF · TaxiBJ | MAE: 22.1 -> 19.5 | Hurt |
| add spatio-temporal circuit to LIF | LIF · KTH (10→40) | PSNR: 20.89 -> 22.63 | Helped |
| add spatio-temporal circuit to PLIF | PLIF · Moving MNIST | MSE: 99.2 -> 48.4 | Helped |
| add spatio-temporal circuit to LM-H | LM-H · Moving MNIST | MSE: 93.2 -> 45.9 | Helped |
| add spatio-temporal circuit to LIF | LIF · CIFAR100, 4 time-steps | Accuracy (%): 63.40 -> 66.64 | Helped |
| remove temporal circuit (ablation) | STC-LIF · Moving MNIST | MSE: 47.0 -> 62.5 | Hurt |
| remove spatial circuit (ablation) | STC-LIF · Moving MNIST | MSE: 47.0 -> 58.3 | Hurt |
| use only self-connection instead of group convolution (ablation) | STC-LIF · Moving MNIST | MSE: 47.0 -> 64.5 | Hurt |
| use global convolution instead of group convolution (ablation) | STC-LIF · Moving MNIST | MSE: 47.0 -> 38.5 | Helped |
Surprising findings
- The vanilla LIF model achieves similar accuracy on shuffled and sequential DVS128 Gesture input, suggesting it mainly learns spatial rather than temporal structure.
- Global convolution gives better Moving MNIST MSE (38.5) than group convolution (47.0) but at roughly 3.5x the parameters.
Limitations
- Gains on TaxiBJ are modest relative to GLIF and TC-LIF on some metrics; STC-LIF MAE is slightly worse than TC-LIF.
- Experiments are limited to a small 4-layer architecture on low-resolution benchmarks, with a single NVIDIA 3090 GPU.
- Gradient truncation through the self-connection circuit is a design choice that decouples output and regulation, with no broader analysis.
Experimental setup
Datasets: Moving MNIST TaxiBJ KTH DVS128 Gesture CIFAR100
Benchmarks: Moving MNIST TaxiBJ KTH DVS128 Gesture CIFAR100
Metrics: MSE MAE SSIM PSNR
Baselines: Vanilla LIF PLIF GLIF TC-LIF LM-H
Builds on: LIF PLIF GLIF TC-LIF LM-H OpenSTL
Closest papers by meaning
- HetSyn: Versatile Timescale Integration in Spiking Neural Networks via Heterogeneous Synapses (NeurIPS 2025 · similarity 0.89)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.88)
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.88)
- SpikingSSMs: Learning Long Sequences with Sparse and Parallel Spiking State Space Models (AAAI 2025 · similarity 0.88)
- Rethinking the Membrane Dynamics and Optimization Objectives of Spiking Neural Networks (NeurIPS 2024 · similarity 0.88)
- TC-LIF: A Two-Compartment Spiking Neuron Model for Long-Term Sequential Modelling (AAAI 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.