2018/05/21 by Drew Linsley, Linsley, Drew, Junkyung Kim +5 · 5 citations
Biochemistry, Genetics and Molecular Biology · Computer Science · Engineering · #Cell Image Analysis Techniques #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Image Processing Techniques and Applications #Medical Image Segmentation Techniques
paper · pdf · doi:10.48550/arxiv.1805.08315
openalex publication_date 2018/05/21 · openalex created_date 2022/08/15 · openalex updated_date 2026/07/28
Progress in deep learning has spawned great successes in many engineering\napplications. As a prime example, convolutional neural networks, a type of\nfeedforward neural networks, are now approaching -- and sometimes even\nsurpassing -- human accuracy on a variety of visual recognition tasks. Here,\nhowever, we show that these neural networks and their recent extensions\nstruggle in recognition tasks where co-dependent visual features must be\ndetected over long spatial ranges. We introduce the horizontal gated-recurrent\nunit (hGRU) to learn intrinsic horizontal connections -- both within and across\nfeature columns. We demonstrate that a single hGRU layer matches or outperforms\nall tested feedforward hierarchical baselines including state-of-the-art\narchitectures which have orders of magnitude more free parameters. We further\ndiscuss the biological plausibility of the hGRU in comparison to anatomical\ndata from the visual cortex as well as human behavioral data on a classic\ncontour detection task.\n