Research map / Architectures and efficiency / Spiking networks
Dendritic Resonate-and-Fire Neuron for Effective and Efficient Long Sequence Modeling
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 the Dendritic Resonate-and-Fire (D-RF) spiking neuron, with multiple frequency-selective dendritic RF branches feeding a soma with a history-dependent adaptive threshold, enabling parallel training and sparse spiking for long sequence modeling.
Problem
Resonate-and-Fire spiking neurons have narrow frequency bandwidth and face a trade-off between sparse spiking (requires reset, costly O(L^2) training) and fast training (no reset, excessive spiking) on long sequence tasks.
Approach
Replace a single RF neuron with multiple dendritic branches, each an RF resonator with its own decay and frequency, combined by importance weights, and add a soma adaptive threshold computed as a causal convolution over past spikes, so the whole neuron remains parallelizable via FFT convolution while producing sparse spikes.
Key results
- Achieves 96.20% accuracy on SHD and 84.30% on S-CIFAR10, with 99.50/98.20 on S/PS-MNIST, outperforming the listed baselines.
- On LRA, reaches 82.88% average accuracy, the best among listed SNN models, with 60.02 on ListOps and 86.52 on Text.
- Achieves 6.1% average spiking rate and 135.8 mJ average energy on LRA, versus 9.3% and 220.6 mJ for SD-TCM (reproduced by the authors); reports up to 147x training acceleration over sequential training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add D-RF instead of PMSN | spiking sequence models · S-CIFAR10 | accuracy: 82.14 -> 84.30 | Helped |
| add D-RF instead of SpikingSSMs | spiking sequence models · LRA average | accuracy: 82.23 -> 82.88 | Mixed |
| add adaptive threshold (ablation) | D-RF · S-CIFAR10 | accuracy: 83.9 -> 84.3 | Helped |
| add adaptive threshold (ablation) | D-RF · LRA ListOps | accuracy: 59.2 -> 60.2 | Helped |
| increase dendrites from n=1 to n=16 (ablation) | D-RF with adaptive threshold · S-CIFAR10 | accuracy: 80.3 -> 85.1 | Helped |
| increase dendrites from n=1 to n=16 (ablation) | D-RF with adaptive threshold · LRA ListOps | accuracy: 55.2 -> 60.3 | Helped |
| add adaptive threshold | D-RF · LRA Image | spiking rate | Helped |
| use D-RF instead of BPTT | long-sequence training · sequence length 32768 | training speedup: 581× | Helped |
| parallelize D-RF training with adaptive threshold | D-RF · Text (4096) | training acceleration: up to 147× | Helped |
Limitations
- Image task on LRA underperforms SpikingSSM, attributed to missing LayerNorm.
- Some SNN baselines (SD-TCM) had to be reproduced by the authors because the original code was not public.
- Performance gap of up to about 3% relative to ANN-based S4 on LRA tasks.
Experimental setup
Datasets: Spiking Heidelberg Digits (SHD) Sequential MNIST Permuted Sequential MNIST Sequential CIFAR10 Long Range Arena (ListOps, Text, Retrieval, Image, Pathfinder)
Benchmarks: SHD S-MNIST PS-MNIST S-CIFAR10 Long Range Arena
Metrics: accuracy spiking rate (%) theoretical energy cost (mJ) training time per epoch speedup
Baselines: LIF ALIF BRF PSN TC-LIF DH-LIF PMSN SPSN SpikingSSM SD-TCM S4 Spiking LMU
Builds on: Resonate-and-Fire neuron PMSN SD-TCM Spiking S4 / SpikingSSM Dendritic heterogeneity (DH-LIF)
Closest papers by meaning
- Balanced Resonate-and-Fire Neurons (ICML 2024 · similarity 0.91)
- 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 0.85)
- TC-LIF: A Two-Compartment Spiking Neuron Model for Long-Term Sequential Modelling (AAAI 2024 · similarity 0.85)
- LongSpike: Fractional Order Spiking State Space Models for Efficient Long Sequence Learning (NeurIPS 2026 · similarity 0.85)
- FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping (NeurIPS 2026 · similarity 0.85)
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.