The study presents a counterintuitive approach to unsupervised perceptual grouping and segmentation in Deep Neural Networks (DNNs). It proposes that these arise due to neural noise, instead of in spite of it. The authors mathematically demonstrate this phenomenon and show that adding noise to a DNN allows it to segment images without any prior training on segmentation labels. The research also introduces the Good Gestalt datasets designed specifically to test perceptual grouping. The study suggests a new explanation for the formation of perceptual grouping and potential benefits of neural noise in the visual system.
Publication date: 28 Sep 2023
Project Page: https://arxiv.org/abs/2309.16515
Paper: https://arxiv.org/pdf/2309.16515
