BROWSE
Detail
Conference Paper
N-MNIST SNN 입력 표현의 정확도-지연 비교
Cited - time in 
Share 
Share
- Issue Date
- 2026-07
- Citation
- 제어·로봇·시스템학회 학술 대회 (ICROS) 2026, pp.1-2
- Publisher
- 제어·로봇·시스템학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- This paper compares three N-MNIST input representations for DVS-to-SNN inference: count-frame, denoised count-frame, and binary event occupancy. All baselines use identical temporal slicing, seeds 3407/3413/3421, 30 training epochs, and the official test split. Count-frame accumulates ON/OFF event counts, denoised count-frame applies Tonic denoise before the same count representation, and binary occupancy stores two-channel ON/OFF occurrence slices. Denoised count-frame achieved the highest test accuracy and lowest end-to-end latency, 89.33 % and 96.15 ms. Binary event occupancy achieved 89.20 %, improving count-frame by 3.49 percentage points and reducing latency by 16.05 ms, with a 1.01 ms gap to denoised count-frame under dense time-slice profiling. Total software memory was similar, 1.66-1.67 MB. The comparison shows how event counting, denoising, and binary occupancy affect accuracy and latency under the same SNN evaluation protocol.
- Keyword
- Spiking neural network, dynamic vision sensor, event-based vision, neuromorphic preprocessing
- KSP Keywords
- End to End(E2E), Evaluation Protocol, Percentage points, Spiking Neural Network, Time Slice, Training Epochs, dynamic vision sensor, end-to-end latency, event-based vision