{"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/what-do-deep-networks-like-to-see","title":"What do Deep Networks Like to See?","arxiv_id":"1803.08337","date":"2018-03-22","proceeding":"CVPR 2018 6","authors":["Sebastian Palacio","Joachim Folz","Jörn Hees","Federico Raue","Damian Borth","Andreas Dengel"],"abstract":"We propose a novel way to measure and understand convolutional neural\nnetworks by quantifying the amount of input signal they let in. To do this, an\nautoencoder (AE) was fine-tuned on gradients from a pre-trained classifier with\nfixed parameters. We compared the reconstructed samples from AEs that were\nfine-tuned on a set of image classifiers (AlexNet, VGG16, ResNet-50, and\nInception~v3) and found substantial differences. The AE learns which aspects of\nthe input space to preserve and which ones to ignore, based on the information\nencoded in the backpropagated gradients. Measuring the changes in accuracy when\nthe signal of one classifier is used by a second one, a relation of total order\nemerges. This order depends directly on each classifier's input signal but it\ndoes not correlate with classification accuracy or network size. Further\nevidence of this phenomenon is provided by measuring the normalized mutual\ninformation between original images and auto-encoded reconstructions from\ndifferent fine-tuned AEs. These findings break new ground in the area of neural\nnetwork understanding, opening a new way to reason, debug, and interpret their\nresults. We present four concrete examples in the literature where observations\ncan now be explained in terms of the input signal that a model uses.","url_abs":"http://arxiv.org/abs/1803.08337v1","url_pdf":"http://arxiv.org/pdf/1803.08337v1.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":"what-do-deep-networks-like-to-see","repo_url":"https://github.com/spalaciob/s2snets-reconstruction","is_official":1,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"image-classification","task_name":"Image Classification"}],"methods":[{"method_slug":"ae","method_name":"AE"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/image-classification-on-imagenet","task":"Image Classification","dataset":"ImageNet","model":"Inception v3","rank_in_archive_order":884,"of":1060,"metrics":{"Top 1 Accuracy":"77.12%"},"uses_additional_data":false}],"syntology":{"syntology_url":null,"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}