{"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/adaptive-neural-networks-for-efficient","title":"Adaptive Neural Networks for Efficient Inference","arxiv_id":"1702.07811","date":"2017-02-25","proceeding":"ICML 2017 8","authors":["Tolga Bolukbasi","Joseph Wang","Ofer Dekel","Venkatesh Saligrama"],"abstract":"We present an approach to adaptively utilize deep neural networks in order to\nreduce the evaluation time on new examples without loss of accuracy. Rather\nthan attempting to redesign or approximate existing networks, we propose two\nschemes that adaptively utilize networks. We first pose an adaptive network\nevaluation scheme, where we learn a system to adaptively choose the components\nof a deep network to be evaluated for each example. By allowing examples\ncorrectly classified using early layers of the system to exit, we avoid the\ncomputational time associated with full evaluation of the network. We extend\nthis to learn a network selection system that adaptively selects the network to\nbe evaluated for each example. We show that computational time can be\ndramatically reduced by exploiting the fact that many examples can be correctly\nclassified using relatively efficient networks and that complex,\ncomputationally costly networks are only necessary for a small fraction of\nexamples. We pose a global objective for learning an adaptive early exit or\nnetwork selection policy and solve it by reducing the policy learning problem\nto a layer-by-layer weighted binary classification problem. Empirically, these\napproaches yield dramatic reductions in computational cost, with up to a 2.8x\nspeedup on state-of-the-art networks from the ImageNet image recognition\nchallenge with minimal (<1%) loss of top5 accuracy.","url_abs":"http://arxiv.org/abs/1702.07811v2","url_pdf":"http://arxiv.org/pdf/1702.07811v2.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":"adaptive-neural-networks-for-efficient","repo_url":"https://github.com/ibodumas/PYREST-ADANET","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"MIT"}},{"paper_slug":"adaptive-neural-networks-for-efficient","repo_url":"https://github.com/NervanaSystems/distiller","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"pytorch","reach":{"status":"gone","observed_at":"2026-09-17","how":"tree_404+repo_404"}}],"tasks":[{"task_slug":"binary-classification","task_name":"Binary Classification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1702.07811","atlas_url":"https://app.syntology.ai/?focus=1702.07811","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}