Backpropagation Techniques in SNN and Application in Image Segmentation

Recent advancements in Spiking Neural Networks (SNNs) have begun to bridge the performance gap with Artificial Neural Networks (ANNs), particularly through the use of approximated gradients. This paper delves into modern backpropagation techniques in SNNs, focusing on surrogate or approximate gradient methods. Traditionally, SNNs have been predominantly applied to image classification challenges, incor-porating fully connected dense layers. Our research extends the application of SNNs to the realm of image segmentation. We propose and implement a modified U-net architecture within an SNN framework, tailored to operate without dense layers, producing segmented images as output. Experiments conducted on datasets such as "EM segmentation 2015" and "Data Science Bowl 2018" demonstrate the efficacy of our approach, yielding results that closely rival those achieved by ANN-based methods. This exploration into SNNs for image segmentation not only broadens the scope of SNN applications but also underscores the potential of these networks in tasks traditionally dominated by ANNs.