Distance-Based Out-of-Distribution Image Detection using Spiking Patterns

As machine learning models become increasingly integral to applications such as autonomous systems, medical imaging, and classification tasks, their performance can degrade when faced with data that differs from the training set. The issue of Out-of-Distribution (OoD) detection is critical for ensuring model reliability in real-world scenarios. Spiking Neural Networks (SNNs), which have a potential to enhance energy efficiency and reduce computational requirements compared to traditional models, have emerged as a promising solution for OoD detection. In this paper, we investigate the potential of SNNs for OoD classification by extracting features from three distinct architectures: spike-Conv, spike-ResNet10, and spike-ResNet18. We explore various approaches for OoD image detection based on distances between spiking patterns and representatives of training spiking patterns. Our experimental results demonstrate the potential of spiking-based feature vectors obtained from residual networks for OoD detection.