Research map / Architectures and efficiency / Spiking networks
HetSyn: Versatile Timescale Integration in Spiking Neural Networks via Heterogeneous Synapses
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 HetSyn, a framework giving each synapse its own time constant (HetSynLIF), shown to subsume LIF, ALIF and neuron-level heterogeneous models and to improve SNN accuracy on SHD, S-MNIST, TiDigits and Ti46 and on pattern generation and delayed match-to-sample tasks.
Problem
Spiking neural networks typically use one membrane time constant per neuron, ignoring the synapse-level heterogeneity seen in biology, which limits multi-timescale temporal processing.
Approach
Replace the neuron-level membrane time constant with synapse-specific decay factors so that each afferent's synaptic current is integrated with its own time constant, with reset modeled as a negative current; vanilla LIF, threshold-adaptive LIF and neuron-level heterogeneous LIF arise as special cases.
Key results
- HetSynLIF reaches 92.36% accuracy on SHD with a two-layer recurrent network, reported as the best among similar-architecture models.
- HetSynLIF reaches 98.93% on S-MNIST, 98.99% on TiDigits and 96.53% on Ti46-Alpha, reported as outperforming listed prior methods.
- In delayed match-to-sample, R-HetSynLIF keeps 100% accuracy with 80% of synaptic time constants masked.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HetSynLIF instead of DH-SFNN | Two-layer recurrent SNN · SHD | accuracy: 92.1 -> 92.36 | Helped |
| Use HetSynLIF instead of DH-SRNN | SRNN · S-MNIST | accuracy: 98.87 -> 98.93 | Helped |
| Use HetSynLIF instead of BPT-SNN | SRNN · TiDigits | accuracy: 98.1 -> 98.99 | Helped |
| Use HetSynLIF instead of RSNN | SRNN · Ti46-Alpha | accuracy: 96.44 -> 96.53 | Helped |
| Add synapse-level heterogeneous time constants rather than using homogeneous LIF dynamics (ablation) | Pattern generation task, FSNN and RSNN · Pattern generation | MSE | Helped |
| Add synapse-level heterogeneous time constants rather than homogeneous LIF dynamics (ablation) | Delayed match-to-sample FSNN · Delayed match-to-sample | accuracy | Helped |
| Mask 80% of synaptic time constants so they are non-trainable (ablation) | R-HetSynLIF · Delayed match-to-sample | accuracy: 80%; 100% | No clear effect |
| Add synapse-level heterogeneous time constants rather than homogeneous LIF dynamics (ablation) | SHD speech recognition RSNN · SHD | accuracy | Helped |
| Use synapse-level heterogeneity rather than neuron-level heterogeneity (ablation) | SHD speech recognition RSNN · SHD | accuracy | Helped |
| Add synapse-level heterogeneous time constants (ablation) | HetSynLIF · SHD under deletion noise | accuracy | Helped |
| Add synapse-level heterogeneous time constants (ablation) | HetSynLIF · SHD under time warp | accuracy | Helped |
Surprising findings
- Threshold adaptation (HomNeuALIF) yields limited gains on SHD, performing less well than neuron-level heterogeneity.
Limitations
- Only a subset of synaptic properties (time constants) is modeled; other synaptic features are not explored.
- Architectures are relatively simple (one- or two-layer networks).
- Learned time-constant distribution is only mildly long-tailed, loosely matching biology.
Experimental setup
Datasets: SHD S-MNIST TiDigits Ti46-Alpha Allen Institute Synaptic Physiology Dataset
Benchmarks: SHD S-MNIST TiDigits Ti46-Alpha pattern generation delayed match-to-sample
Metrics: accuracy mean squared error normalized distance success rate
Baselines: DH-SFNN DH-SRNN SRNN NeuHet-SRNN LSTM LSNN AHP-SNN HomNeuLIF HomNeuALIF HetNeuLIF
Builds on: LIF neuron model PLIF ALIF (adaptive LIF) surrogate gradient training (STBP) neuron heterogeneity (Perez-Nieves et al.)
Closest papers by meaning
- Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks (ICML 2024 · similarity 0.89)
- TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting (ICLR 2025 · similarity 0.88)
- TC-LIF: A Two-Compartment Spiking Neuron Model for Long-Term Sequential Modelling (AAAI 2024 · similarity 0.88)
- SpikingSSMs: Learning Long Sequences with Sparse and Parallel Spiking State Space Models (AAAI 2025 · similarity 0.87)
- P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.87)
- Training High Performance Spiking Neural Network by Temporal Model Calibration (ICML 2025 · similarity 0.87)
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.