Research map / Time series / Long-term forecasting

TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting

ICLR 2025 · Time series › Long-term forecasting · approach: Spatiotemporal forecasting models

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 TS-LIF, a dual-compartment spiking neuron with dendritic and somatic branches, direct somatic current injection, and dendritic spiking, evaluated as a drop-in for TCN, GRU, and Transformer SNN forecasters on four traffic/energy benchmarks.

Problem

Standard LIF spiking neurons struggle to capture long-term dependencies and multi-scale temporal dynamics, limiting SNN-based time series forecasting.

Approach

A two-compartment LIF neuron where the dendritic compartment acts as a low-pass filter and the somatic compartment as a high-pass filter, with direct input current injected into the soma and spikes emitted from both compartments, mixed by a learnable weighted sum.

Key results

• TS-former achieves average R^2 0.783 and RSE 0.441 across the four datasets, with the best average rank (3.5) among the compared models.
• On Electricity with missing-input ratios up to 80%, TS-GRU at prediction length 96 keeps R^2 0.922 versus 0.760 for Spike-GRU in the robustness table.
• TS-TCN on Electricity (prediction length 96) uses 0.19 G OPs and 0.25 mJ energy, a 60.93% reduction versus the dense TCN's 0.64 mJ, though TS-TCN has a higher OPs count than the dense TCN's 0.14 G.

Measured effects reported in the paper

Surprising findings

• The paper's own energy table shows TS-TCN consumes more operations (0.19 G) than the dense TCN (0.14 G), with energy savings depending on the SNN accounting.

Limitations

• Stability and frequency analysis ignore spike generation and reset, so the theory applies only to the linearized homogeneous part.
• The frequency-separation example uses hand-picked parameters rather than learned ones.
• Energy estimates are theoretical rather than measured on neuromorphic hardware.

Experimental setup

Datasets: Metr-la Pems-bay Solar Electricity

Benchmarks: Metr-la Pems-bay Solar Electricity

Metrics: R^2 RSE

Baselines: ARIMA GP TCN Spike-TCN GRU Spike-GRU Spike-RNN Autoformer iTransformer iSpikformer TC-LIF LM-H

Builds on: Leaky Integrate-and-Fire TC-LIF LM-H iSpikformer iTransformer

Closest papers by meaning

• TC-LIF: A Two-Compartment Spiking Neuron Model for Long-Term Sequential Modelling (AAAI 2024 · similarity 0.92)
• Efficient and Effective Time-Series Forecasting with Spiking Neural Networks (ICML 2024 · similarity 0.90)
• TS-SNN: Temporal Shift Module for Spiking Neural Networks (ICML 2025 · similarity 0.89)
• CLIF: Complementary Leaky Integrate-and-Fire Neuron for Spiking Neural Networks (ICML 2024 · similarity 0.88)
• P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.88)
• HetSyn: Versatile Timescale Integration in Spiking Neural Networks via Heterogeneous Synapses (NeurIPS 2025 · similarity 0.88)