| name | fits-interpretable-spiking-neuron |
| description | FiTS (Frequency Selectivity and Temporal Shaping) interpretable spiking neuron model. Factorizes temporal computation into Frequency Selectivity and Temporal Shaping within each neuron. Use when: spiking neural networks, interpretable neurons, temporal processing, frequency selectivity, temporal shaping, LIF neuron improvement, auditory processing, neuron-level interpretability, group-delay modulation, subthreshold magnitude response, feedforward SNNs, spike timing. |
FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping
Paper Reference
- Title: FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping
- arXiv: 2605.13071
- Authors: Jongmin Choi, Joon Son Chung
- Date: 2026-05-19
- Category: cs.NE (Neural and Evolutionary Computing)
Core Innovation
FiTS introduces a new spiking neuron model that factorizes temporal computation within each neuron into two independent modules:
- Frequency Selectivity (FS): Parameterizes each neuron's target frequency as the maximizer of its subthreshold magnitude response
- Temporal Shaping (TS): Reshapes when frequency components contribute to membrane voltage accumulation through group-delay modulation
Architecture
FS Module - Frequency Selectivity
- Each neuron learns a target frequency parameter
- Subthreshold membrane dynamics are tuned to maximize response at the target frequency
- Provides frequency-selective filtering at the single-neuron level
- Eliminates need for network-level delays or recurrence for temporal modeling
TS Module - Temporal Shaping
- Controls when frequency components contribute to membrane voltage accumulation
- Uses group-delay modulation to shift temporal alignment of frequency components
- Enables phase alignment and temporal feature extraction at neuron level
Key Properties
Interpretability
- Target frequencies provide interpretable neuron-level summaries of frequency organization
- Group-delay shifts reveal timing organization learned within the network
- Unlike black-box SNN neurons, FiTS neurons have physically meaningful parameters
Performance
- Improves over plain LIF baseline on auditory benchmarks
- Competitive with strong temporal SNN baselines (with recurrence/delays)
- Works in simple feedforward SNNs without recurrence or network-level delays
Comparison to LIF
| Feature | LIF | FiTS |
|---|
| Frequency selectivity | No (fixed time constant) | Yes (learnable target frequency) |
| Temporal shaping | No | Yes (group-delay modulation) |
| Interpretability | Limited | High (frequency + timing parameters) |
| Needs recurrence for temporal | Often | No |
Mathematical Foundation
Subthreshold Response
The FiTS neuron's subthreshold dynamics are designed such that:
$$H(\omega) = \frac{1}{1 + j\omega\tau - \text{FS terms}}$$
The target frequency $\omega^*$ is learned as the maximizer of $|H(\omega)|$.
Group Delay
$$\tau_g(\omega) = -\frac{d}{d\omega}\arg H(\omega)$$
The TS module modulates $\tau_g$ to control temporal alignment.
Use Cases
- Auditory processing: Where frequency selectivity and timing are central
- Temporal pattern recognition: Event-driven temporal processing
- Interpretable SNNs: When neuron-level understanding is required
- Energy-efficient temporal modeling: Feedforward SNNs without recurrence overhead
Activation Keywords
FiTS, frequency selectivity, temporal shaping, interpretable spiking neurons, LIF neuron, group delay, subthreshold response, auditory SNN, feedforward SNN, temporal processing, spiking neuron design, neuron specialization
