Research map / Architectures and efficiency / Spiking networks

FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping

NeurIPS 2026 · Architectures and efficiency › Spiking networks · approach: Spiking transformers · 0 citations (OpenAlex)

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 FiTS, a spiking neuron whose adaptive-LIF subthreshold response has a learnable target frequency (closed-form inverse map) plus an all-pass/lambda-mixing group-delay module, improving feedforward SNNs on SHD, SSC and GSC.

Problem

Individual spiking neurons' frequency preference and response timing are typically implicit in learned coefficients or network-level mechanisms; it is unclear how a single neuron should specialize its temporal role within a feedforward SNN.

Approach

Parameterize each neuron's target frequency as the maximizer of its subthreshold magnitude response (FS module, with closed-form inverse to adaptation strength) and reshape pre-spike group delay with an all-pass cascade mixed via learnable lambda weights (TS module).

Key results

• FiTS reaches 95.31 ± 0.21% on SHD (best-test protocol) and 94.38 ± 0.12% under the 20% validation-split protocol with 0.04M parameters.
• FiTS reaches 78.23 ± 0.16% on SSC, improving over plain LIF (69.17 ± 0.18%) at width 512 with 0.35M parameters.
• FiTS reaches 94.48 ± 0.12% on GSC versus 80.22 ± 0.16% for plain LIF.

Surprising findings

• Adaptation with zero-initialized target frequency shows no consistent improvement over plain LIF; gains come mainly from log-scale nonzero target-frequency initialization.
• TS module reduces linear-layer spike cost but raises neuron-internal computation, so total estimated energy increases from 8.9 to 20.3 uJ at width 512.

Limitations

• Evaluation focuses on auditory benchmarks; sMNIST and DVS Gesture only in appendix.
• TS module adds neuron-internal computation, increasing total theoretical energy cost.
• Does not combine with recurrence or network-level delays; on SSC remains below DH-SFNN and C-SiLIF.

Experimental setup

Datasets: SHD SSC GSC v0.02

Benchmarks: SHD SSC GSC

Metrics: top-1 accuracy parameter count theoretical energy (uJ)

Baselines: RadLIF S5-RF SE-adLIF DCLS-Delays DelRec TC-LIF cAdLIF DH-SFNN C-SiLIF SiLIF

Builds on: adaptive LIF surrogate gradient resonate-and-fire SE-adLIF cAdLIF

Closest papers by meaning

• Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness (ICLR 2025 · similarity 0.89)
• Learning Delays in Spiking Neural Networks using Dilated Convolutions with Learnable Spacings (ICLR 2024 · similarity 0.89)
• Spiking Neural Networks Need High-Frequency Information (NeurIPS 2025 · similarity 0.88)
• FSTA-SNN:Frequency-Based Spatial-Temporal Attention Module for Spiking Neural Networks (AAAI 2025 · similarity 0.88)
• Balanced Resonate-and-Fire Neurons (ICML 2024 · similarity 0.87)
• From Cortical Synchronous Rhythm to Brain Inspired Learning Mechanism: An Oscillatory Spiking Neural Network with Time-Delayed Coordination (NeurIPS 2026 · similarity 0.87)