Simulations show time‑varying signals reduce spike timing jitter in a CMOS analog neuron
This paper reports that a CMOS analog spiking neuron can reproduce a surprising feature seen in biology: when driven by a time‑varying input, its output spikes are much more tightly timed than when driven by a steady current. The authors used an industrial, physics‑based SPICE transient noise simulation framework that includes standard foundry transistor models and the physical random noise those devices produce. Their main observation is that time‑varying inputs can trigger spikes in a controlled way and cut the spike‑time jitter that builds up under constant drive.
The neuron model studied is an ultra‑low‑power CMOS circuit based on a simplified Morris–Lecar model previously described by Sourikopoulos et al. The designers kept the transistor sizes and supply voltage from that design and added the real device noise present in the transistor models, especially thermal noise. They then replayed two kinds of inputs analogous to classic neuroscience experiments: a constant (supra‑threshold) current and a time‑varying current made of a steady component plus a frozen noisy waveform. Using many simulated trials, they compared how variable the spike times were under the two inputs.
At a high level, the difference comes from two operating regimes. With a strong constant input the neuron runs as a self‑sustained oscillator: it emits spikes repeatedly and small random fluctuations in each cycle add up in time, producing phase noise or jitter. By contrast, with a time‑varying input the neuron behaves as an excitable system: it sits quietly until a transient positive input pushes it into a single spike (a limit cycle) and then it returns to silence. That event‑like triggering avoids the cumulative phase error and keeps spike times aligned across runs.
This result matters for neuromorphic hardware and for how we think about spike timing in artificial neurons. It supports the idea that event‑based or timing‑based coding can be more reliable than rate coding when real device noise is present. The finding also highlights that device thermal noise matters for timing precision in low‑power analog circuits, and that input shaping can be an effective way to improve reliability without changing the circuit layout.