FiTS spiking neurons improve temporal processing with frequency selectivity
New neuron model decodes sound like the brain, one frequency at a time
Deep Dive
Choi and Chung introduce FiTS, a spiking neuron that separates temporal computation into frequency selectivity and temporal shaping. On auditory benchmarks, feedforward FiTS networks improve over plain LIF neurons without recurrence or delays, remaining competitive with strong temporal SNN baselines. The learned frequency targets and group-delay shifts offer interpretable neuron-level summaries of network organization.
Key Points
- FiTS factorizes each spiking neuron into Frequency Selectivity (FS) and Temporal Shaping (TS) modules.
- On auditory benchmarks, FiTS outperforms LIF baselines and matches complex temporal SNNs without recurrence or delays.
- Learned target frequencies and group-delay shifts provide interpretable, neuron-level summaries of network organization.
Why It Matters
Interpretable SNNs bring transparency to event-driven AI, crucial for high-stakes audio and time-series applications.