{"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/net2vec-quantifying-and-explaining-how","title":"Net2Vec: Quantifying and Explaining how Concepts are Encoded by Filters in Deep Neural Networks","arxiv_id":"1801.03454","date":"2018-01-10","proceeding":"CVPR 2018 6","authors":["Ruth Fong","Andrea Vedaldi"],"abstract":"In an effort to understand the meaning of the intermediate representations\ncaptured by deep networks, recent papers have tried to associate specific\nsemantic concepts to individual neural network filter responses, where\ninteresting correlations are often found, largely by focusing on extremal\nfilter responses. In this paper, we show that this approach can favor\neasy-to-interpret cases that are not necessarily representative of the average\nbehavior of a representation.\n  A more realistic but harder-to-study hypothesis is that semantic\nrepresentations are distributed, and thus filters must be studied in\nconjunction. In order to investigate this idea while enabling systematic\nvisualization and quantification of multiple filter responses, we introduce the\nNet2Vec framework, in which semantic concepts are mapped to vectorial\nembeddings based on corresponding filter responses. By studying such\nembeddings, we are able to show that 1., in most cases, multiple filters are\nrequired to code for a concept, that 2., often filters are not concept specific\nand help encode multiple concepts, and that 3., compared to single filter\nactivations, filter embeddings are able to better characterize the meaning of a\nrepresentation and its relationship to other concepts.","url_abs":"http://arxiv.org/abs/1801.03454v2","url_pdf":"http://arxiv.org/pdf/1801.03454v2.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":"net2vec-quantifying-and-explaining-how","repo_url":"https://github.com/ruthcfong/net2vec","is_official":1,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"unanswered"}}],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1801.03454","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}