{"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-makes-imagenet-good-for-transfer","title":"What makes ImageNet good for transfer learning?","arxiv_id":"1608.08614","date":"2016-08-30","proceeding":null,"authors":["Minyoung Huh","Pulkit Agrawal","Alexei A. Efros"],"abstract":"The tremendous success of ImageNet-trained deep features on a wide range of\ntransfer tasks begs the question: what are the properties of the ImageNet\ndataset that are critical for learning good, general-purpose features? This\nwork provides an empirical investigation of various facets of this question: Is\nmore pre-training data always better? How does feature quality depend on the\nnumber of training examples per class? Does adding more object classes improve\nperformance? For the same data budget, how should the data be split into\nclasses? Is fine-grained recognition necessary for learning good features?\nGiven the same number of training classes, is it better to have coarse classes\nor fine-grained classes? Which is better: more classes or more examples per\nclass? To answer these and related questions, we pre-trained CNN features on\nvarious subsets of the ImageNet dataset and evaluated transfer performance on\nPASCAL detection, PASCAL action classification, and SUN scene classification\ntasks. Our overall findings suggest that most changes in the choice of\npre-training data long thought to be critical do not significantly affect\ntransfer performance.? Given the same number of training classes, is it better\nto have coarse classes or fine-grained classes? Which is better: more classes\nor more examples per class?","url_abs":"http://arxiv.org/abs/1608.08614v2","url_pdf":"http://arxiv.org/pdf/1608.08614v2.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-makes-imagenet-good-for-transfer","repo_url":"https://github.com/minyoungg/wmigftl","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok"}}],"tasks":[{"task_slug":"action-classification","task_name":"Action Classification"},{"task_slug":"classification","task_name":"General Classification"},{"task_slug":"scene-classification","task_name":"Scene Classification"},{"task_slug":"transfer-learning","task_name":"Transfer Learning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":"https://app.syntology.ai/?focus=1608.08614","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}