PMSM: Nearly lossless ANN-to-SNN conversion in just one timestep
First-ever ultra-low-latency conversion without accuracy loss on image and neuromorphic datasets.
A new paper from a team of researchers presents Polarity Multi-Spike Mapping (PMSM), a method for converting artificial neural networks (ANNs) to spiking neural networks (SNNs) that is nearly lossless and works at the theoretical minimum of one timestep. Traditional ANN-to-SNN conversion methods suffer from accuracy loss or require multiple timesteps because quantization layers discard valuable negative-value information after batch normalization. The PMSM framework addresses this by mapping both positive and negative activations — a “polarity multi-spike” approach — and introducing a hyperparameter initialization strategy tuned via information entropy analysis.
The team tested PMSM on six image and neuromorphic datasets, including CIFAR and event-based benchmarks, and found it achieves accuracy comparable to the original ANN in a single timestep. Remarkably, PMSM even outperforms state-of-the-art direct training methods on several tasks, despite operating at ultra-low latency. This breakthrough makes SNNs more practical for real-time, energy-efficient applications, as event-driven spike computation can now be deployed without the typical latency trade-off.
- PMSM enables nearly lossless ANN-to-SNN conversion at the first timestep, the theoretical minimum latency.
- Framework recovers negative-value information lost in traditional quantization after batch normalization.
- Outperforms direct training methods on multiple benchmarks while using only ultra-low-latency spike computation.
Why It Matters
PMSM makes low-latency, energy-efficient SNNs practical without accuracy trade-offs, accelerating edge-AI and neuromorphic computing adoption.