{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/multigrid-neural-architectures","title":"Multigrid Neural Architectures","arxiv_id":"1611.07661","date":"2016-11-23","proceeding":"CVPR 2017 7","authors":["Tsung-Wei Ke","Michael Maire","Stella X. Yu"],"abstract":"We propose a multigrid extension of convolutional neural networks (CNNs).\nRather than manipulating representations living on a single spatial grid, our\nnetwork layers operate across scale space, on a pyramid of grids. They consume\nmultigrid inputs and produce multigrid outputs; convolutional filters\nthemselves have both within-scale and cross-scale extent. This aspect is\ndistinct from simple multiscale designs, which only process the input at\ndifferent scales. Viewed in terms of information flow, a multigrid network\npasses messages across a spatial pyramid. As a consequence, receptive field\nsize grows exponentially with depth, facilitating rapid integration of context.\nMost critically, multigrid structure enables networks to learn internal\nattention and dynamic routing mechanisms, and use them to accomplish tasks on\nwhich modern CNNs fail.\n  Experiments demonstrate wide-ranging performance advantages of multigrid. On\nCIFAR and ImageNet classification tasks, flipping from a single grid to\nmultigrid within the standard CNN paradigm improves accuracy, while being\ncompute and parameter efficient. Multigrid is independent of other\narchitectural choices; we show synergy in combination with residual\nconnections. Multigrid yields dramatic improvement on a synthetic semantic\nsegmentation dataset. Most strikingly, relatively shallow multigrid networks\ncan learn to directly perform spatial transformation tasks, where, in contrast,\ncurrent CNNs fail. Together, our results suggest that continuous evolution of\nfeatures on a multigrid pyramid is a more powerful alternative to existing CNN\ndesigns on a flat grid.","url_abs":"http://arxiv.org/abs/1611.07661v2","url_pdf":"http://arxiv.org/pdf/1611.07661v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"multigrid-neural-architectures","repo_url":"https://github.com/buttomnutstoast/Multigrid-Neural-Architectures","is_official":1,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"torch","reach":null}],"tasks":[{"task_slug":"image-classification","task_name":"Image Classification"},{"task_slug":"semantic-segmentation","task_name":"Semantic Segmentation"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}