Research map / Architectures and efficiency / Spiking networks
TC-LIF: A Two-Compartment Spiking Neuron Model for Long-Term Sequential Modelling
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 TC-LIF, a two-compartment spiking neuron with somatic and dendritic membrane potentials, to improve long-term temporal credit assignment in SNNs, evaluated on S-MNIST, PS-MNIST, GSC, SHD and SSC.
Problem
Spiking neural networks struggle to establish long-term temporal dependencies between distant cues because error gradients vanish over long time spans in single-compartment neurons such as LIF.
Approach
A biologically inspired two-compartment LIF neuron derived from the Pinsky-Rinzel model, with a dendritic compartment that retains long-term memory, coupled to the soma via opposite-signed learnable coupling coefficients (beta1 in (-1,0), beta2 in (0,1)) and no membrane decay, so gradient propagation over time is improved.
Key results
- TC-LIF reaches 99.20% test accuracy on S-MNIST with a recurrent network, reported as the best SNN result on that dataset.
- On SHD, TC-LIF (recurrent) reaches 88.91% accuracy versus 84.40% for SRNN+ALIF; on GSC recurrent it reaches 94.84% versus 92.10% for SRNN+ALIF.
- Empirical energy cost is 28.2 nJ for TC-LIF versus 23.8 nJ for LIF and 2,834.7 nJ for LSTM, i.e. more than 100x savings over LSTM; TC-LIF converges in about 25 epochs versus about 100 for LIF on feedforward networks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace LIF neurons with TC-LIF neurons | feedforward SNN · S-MNIST | accuracy (%): 94.80/95.27 -> 96.46/97.35 | Helped |
| replace SRNN+ReLU with TC-LIF neurons | recurrent SNN · S-MNIST | accuracy (%): 98.99/ - -> 98.79/99.20 | Mixed |
| replace LIF neurons with TC-LIF neurons | feedforward SNN · PS-MNIST | accuracy (%): 11.30/10.00 -> 80.89/83.98 | Helped |
| replace SRNN+ALIF with TC-LIF neurons | recurrent SNN · PS-MNIST | accuracy (%): 91.00/94.30 -> 92.69/95.36 | Helped |
| replace rate-based SNN with TC-LIF neurons | feedforward SNN · GSC | accuracy (%): 75.20 -> 91.35 | Helped |
| replace SRNN+ALIF with TC-LIF neurons | recurrent SNN · GSC | accuracy (%): 92.10 -> 94.84 | Helped |
| replace feed-forward SNN neurons with TC-LIF neurons | feedforward SNN · SHD | accuracy (%): 48.60 -> 83.08 | Helped |
| replace SRNN+ALIF with TC-LIF neurons | recurrent SNN · SHD | accuracy (%): 84.40 -> 88.91 | Helped |
| replace feed-forward SNN neurons with TC-LIF neurons | feedforward SNN · SSC | accuracy (%): 38.50 -> 63.46 | Helped |
| replace Heterogeneous SRNN with TC-LIF neurons | recurrent SNN · SSC | accuracy (%): 57.3 -> 61.09 | Helped |
| use TC-LIF instead of LIF neurons | recurrent SNN · S-MNIST | empirical energy cost (nJ): 23.8 -> 28.2 | Hurt |
| initialize beta1, beta2 to (-0.2, 0.8) (ablation) | TC-LIF · S-MNIST | test accuracy (%): 99.20 -> 89.36 | Hurt |
Surprising findings
- Single-compartment LIF baselines reach only 11.30% on PS-MNIST feedforward, and the authors report that initializing the coupling coefficients in other quadrants fails to converge.
- The infinite norm of the Jacobian between adjacent time steps stays slightly above 1 after training, so the authors accept mild gradient explosion risk rather than eliminating it.
Limitations
- Only two-compartment models are studied; generalization to more compartments is left open.
- Gradient norm analysis shows the infinite norm exceeds 1 throughout, so exploding gradients are possible; stability is only empirically observed.
- Results on some tasks rely on reproduced baselines marked with asterisk and on hyperparameters tuned per dataset.
Experimental setup
Datasets: S-MNIST PS-MNIST Google Speech Commands v2 Spiking Heidelberg Digits Spiking Google Speech Commands
Benchmarks: S-MNIST PS-MNIST GSC SHD SSC
Metrics: test accuracy parameter count energy cost (nJ) training convergence epochs
Baselines: LIF GLIF PLIF LTMD LSTM SRNN+ReLU LSNN AHP SRNN+ALIF Heterogeneous SRNN
Builds on: Leaky Integrate-and-Fire Pinsky-Rinzel model surrogate gradient LSNN GLIF PLIF
Closest papers by meaning
- TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting (ICLR 2025 · similarity 0.92)
- P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.90)
- CLIF: Complementary Leaky Integrate-and-Fire Neuron for Spiking Neural Networks (ICML 2024 · similarity 0.89)
- Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks (ICML 2024 · similarity 0.88)
- The Expressive Leaky Memory Neuron: an Efficient and Expressive Phenomenological Neuron Model Can Solve Long-Horizon Tasks. (ICLR 2024 · similarity 0.88)
- Neuronal Self-Adaptation Enhances Capacity and Robustness of Representation in Spiking Neural Networks (NeurIPS 2026 · 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.