Out-of-Distribution Image Detection with Residual Spiking Neural Network Architectures

In recent years, the rapid advancements in machine learning models have led to significant improvements across various domains such as medical applications, autonomous systems, and image classification. However, real-world conditions often deviate from the controlled environments in which models are trained, leading to performance degradation. Out-of-distribution (OoD) detection has emerged as a critical task to address this issue, enabling models to identify when input data diverges from the training distribution. Spiking Neural Networks (SNNs), known for their energy efficiency and lower computational complexity compared to traditional models, have gained attention for their potential in OoD detection tasks. This paper explores the effectiveness of SNNs in OoD detection by introducing three distinct architectures: spike-Conv, spike-ResNet9, and spike-ResNet10, trained on widely used datasets. We present a comparative analysis of these architectures in terms of their accuracy and robustness in various OoD detection environments. The experimental results demonstrate that the proposed spiking network architectures can effectively handle OoD detection with satisfcatory accuracy.