Event-driven spike encoders translate analog input signals into sparse spike trains, enabling ultra-low-power neuromorphic analytics. We compare five spike encoders applied to electroencephalogram (EEG) traces: fixed-threshold asynchronous Delta modulation (ADM), adaptive-threshold ADM (ADM-A), first- and second-order asynchronous Sigma-Delta modulation (ASDM-1, ASDM-2) and a multi-level Level-Crossing scheme (LCS). The comparative study aims to identify the most suitable candidate(s) for downstream Spiking Neural Network (SNN) processing. The test dataset comprises eyes-closed (EC) and eyes-open (EO) resting-state 64-channels EEG from 109 subjects, segmented into 4 s, 6 s, and 8 s epochs (ten windows for each subject). Metrics reported are SNR, R2, spike rate, RMSE, band-power error, spike train entropy and power consumption estimation of theoretical circuits associated with the methods. The encoders exhibit clear trade-offs between fidelity, sparsity, and energy. ASDM-2 achieves the best overall fidelity (about 21 dB SNR) at a moderate spike rate (about 360 spikes s−1), whereas ADM-A provides the second-highest fidelity (about 19 dB SNR) at a higher spike rate (about 590 spikes s−1). Behavioral power estimates in a 45nm CMOS setting indicate that ASDM-2 has lower dynamic power and lower energy per spike than ADM-A, while requiring about 40% higher static power because of its additional integration stages. Overall, the proposed benchmarking framework provides a principled basis for selecting spike encoders in mixed-signal neuromorphic EEG front-ends and motivates future studies in which different encoders are coupled with SNNs to evaluate task-level accuracy and energy per inference.
Benchmarking Event-Driven Spike Encoders on Resting EEG: A Comparative Study
Collu R.Secondo
;Raffo L.Penultimo
;Barbaro M.Ultimo
2026-01-01
Abstract
Event-driven spike encoders translate analog input signals into sparse spike trains, enabling ultra-low-power neuromorphic analytics. We compare five spike encoders applied to electroencephalogram (EEG) traces: fixed-threshold asynchronous Delta modulation (ADM), adaptive-threshold ADM (ADM-A), first- and second-order asynchronous Sigma-Delta modulation (ASDM-1, ASDM-2) and a multi-level Level-Crossing scheme (LCS). The comparative study aims to identify the most suitable candidate(s) for downstream Spiking Neural Network (SNN) processing. The test dataset comprises eyes-closed (EC) and eyes-open (EO) resting-state 64-channels EEG from 109 subjects, segmented into 4 s, 6 s, and 8 s epochs (ten windows for each subject). Metrics reported are SNR, R2, spike rate, RMSE, band-power error, spike train entropy and power consumption estimation of theoretical circuits associated with the methods. The encoders exhibit clear trade-offs between fidelity, sparsity, and energy. ASDM-2 achieves the best overall fidelity (about 21 dB SNR) at a moderate spike rate (about 360 spikes s−1), whereas ADM-A provides the second-highest fidelity (about 19 dB SNR) at a higher spike rate (about 590 spikes s−1). Behavioral power estimates in a 45nm CMOS setting indicate that ASDM-2 has lower dynamic power and lower energy per spike than ADM-A, while requiring about 40% higher static power because of its additional integration stages. Overall, the proposed benchmarking framework provides a principled basis for selecting spike encoders in mixed-signal neuromorphic EEG front-ends and motivates future studies in which different encoders are coupled with SNNs to evaluate task-level accuracy and energy per inference.I metadati presenti in IRIS UNICA sono rilasciati con licenza Creative Commons CC0 1.0 Universal, mentre i file delle pubblicazioni sono protetti da diritto d'autore, salvo diversa indicazione.



