{"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/s3pool-pooling-with-stochastic-spatial","title":"S3Pool: Pooling with Stochastic Spatial Sampling","arxiv_id":"1611.05138","date":"2016-11-16","proceeding":"CVPR 2017 7","authors":["Shuangfei Zhai","Hui Wu","Abhishek Kumar","Yu Cheng","Yongxi Lu","Zhongfei Zhang","Rogerio Feris"],"abstract":"Feature pooling layers (e.g., max pooling) in convolutional neural networks\n(CNNs) serve the dual purpose of providing increasingly abstract\nrepresentations as well as yielding computational savings in subsequent\nconvolutional layers. We view the pooling operation in CNNs as a two-step\nprocedure: first, a pooling window (e.g., $2\\times 2$) slides over the feature\nmap with stride one which leaves the spatial resolution intact, and second,\ndownsampling is performed by selecting one pixel from each non-overlapping\npooling window in an often uniform and deterministic (e.g., top-left) manner.\nOur starting point in this work is the observation that this regularly spaced\ndownsampling arising from non-overlapping windows, although intuitive from a\nsignal processing perspective (which has the goal of signal reconstruction), is\nnot necessarily optimal for \\emph{learning} (where the goal is to generalize).\nWe study this aspect and propose a novel pooling strategy with stochastic\nspatial sampling (S3Pool), where the regular downsampling is replaced by a more\ngeneral stochastic version. We observe that this general stochasticity acts as\na strong regularizer, and can also be seen as doing implicit data augmentation\nby introducing distortions in the feature maps. We further introduce a\nmechanism to control the amount of distortion to suit different datasets and\narchitectures. To demonstrate the effectiveness of the proposed approach, we\nperform extensive experiments on several popular image classification\nbenchmarks, observing excellent improvements over baseline models. Experimental\ncode is available at https://github.com/Shuangfei/s3pool.","url_abs":"http://arxiv.org/abs/1611.05138v1","url_pdf":"http://arxiv.org/pdf/1611.05138v1.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":"s3pool-pooling-with-stochastic-spatial","repo_url":"https://github.com/Shuangfei/s3pool","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"none","reach":{"status":"ok"}},{"paper_slug":"s3pool-pooling-with-stochastic-spatial","repo_url":"https://github.com/edgarmedina1801/S3Pool","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}},{"paper_slug":"s3pool-pooling-with-stochastic-spatial","repo_url":"https://github.com/emedinac/S3Pool","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}},{"paper_slug":"s3pool-pooling-with-stochastic-spatial","repo_url":"https://github.com/gan3sh500/custom-pooling","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"data-augmentation","task_name":"Data Augmentation"},{"task_slug":"image-classification","task_name":"Image Classification"},{"task_slug":"image-classification","task_name":"image-classification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1611.05138","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}