모델 잡음 강건성 향상을 위한 Spiking Neuron 적용 가능성 고찰

초록

With increasing interest in real-time video analytics, efforts to achieve this by deploying models on edge devices continue to grow. Periodic retraining methods for lightweight models deployed in edge computing environments have been proposed as an effective approach to mitigate data drift problems. However, if the Golden Model, which generates labels for retraining, produces inaccurate labels, retraining performances of lightweight models might degrade. In this study, spiking neurons were integrated into the Golden Model to enable noise-robust labeling. To evaluate this approach, we assessed the classification accuracy of noisy data using the MNIST, Neuromorphic-MNIST, and DVSGesture datasets. Our results demonstrated that the proposed method achieved classification accuracy comparable to existing approaches, even without prior training on noise.

키워드