FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping

Heading to NeurIPS 2026

FiTS has been accepted to NeurIPS 2026 as a Spotlight paper, and as the first author I will be in Sydney in early December to present it at the poster session.

Spotlight presentation

At a glance

• Temporal computation, inside each neuron. FiTS is a spiking neuron that factorizes the temporal computation within each neuron into Frequency Selectivity (FS) and Temporal Shaping (TS).
• No extra network machinery. On auditory benchmarks, FiTS consistently improves over a plain Leaky Integrate-and-Fire (LIF) baseline in simple feedforward SNNs, without recurrence or network-level delays, while remaining competitive with strong temporal SNN baselines.
• Interpretable by design. The learned target frequencies and group-delay shifts summarize, neuron by neuron, the frequency and timing organization the network has learned.

Motivation

Spiking Neural Networks (SNNs) are a promising framework for event-driven temporal processing. Prior work has improved their temporal modeling through richer neuron dynamics and network-level mechanisms such as recurrence and delays. What remains unclear is how individual spiking neurons should specialize within a network.

How should a single spiking neuron specialize in time: which frequency it responds to, and when those frequency components contribute to its membrane voltage?

Key idea: frequency selectivity and temporal shaping

1. 1

   Frequency Selectivity (FS)

   Parameterizes each neuron's target frequency as the maximizer of its subthreshold magnitude response.

2. 2

   Temporal Shaping (TS)

   Reshapes when frequency components contribute to membrane voltage accumulation, through group-delay modulation.

Key findings

We evaluate FiTS on auditory benchmarks, where frequency selectivity and timing are central to the structure of the input. In simple feedforward SNNs without recurrence or network-level delays, FiTS consistently improves over the LIF baseline while remaining competitive with strong temporal SNN baselines.

Beyond accuracy, the learned target frequencies and group-delay shifts provide interpretable, neuron-level summaries of how the network organizes frequency and timing.

Looking ahead

If you are attending, please come say hello. I would love to talk about spiking neurons, frequency selectivity, and what biology can still teach us about network design. After the conference, I will update this page with questions from the session, key discussion points, and personal takeaways. See you in Sydney!