{"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/deep-learning-improved-by-biological","title":"Deep learning improved by biological activation functions","arxiv_id":"1804.11237","date":"2018-03-19","proceeding":null,"authors":["Gardave S Bhumbra"],"abstract":"`Biologically inspired' activation functions, such as the logistic sigmoid,\nhave been instrumental in the historical advancement of machine learning.\nHowever in the field of deep learning, they have been largely displaced by\nrectified linear units (ReLU) or similar functions, such as its exponential\nlinear unit (ELU) variant, to mitigate the effects of vanishing gradients\nassociated with error back-propagation. The logistic sigmoid however does not\nrepresent the true input-output relation in neuronal cells under physiological\nconditions. Here, bionodal root unit (BRU) activation functions are introduced,\nexhibiting input-output non-linearities that are substantially more\nbiologically plausible since their functional form is based on known\nbiophysical properties of neuronal cells.\n  In order to evaluate the learning performance of BRU activations, deep\nnetworks are constructed with identical architectures except differing in their\ntransfer functions (ReLU, ELU, and BRU). Multilayer perceptrons, stacked\nauto-encoders, and convolutional networks are used to test supervised and\nunsupervised learning based on the MNIST and CIFAR-10/100 datasets. Comparisons\nof learning performance, quantified using loss and error measurements,\ndemonstrate that bionodal networks both train faster than their ReLU and ELU\ncounterparts and result in the best generalised models even in the absence of\nformal regularisation. These results therefore suggest that revisiting the\ndetailed properties of biological neurones and their circuitry might prove\ninvaluable in the field of deep learning for the future.","url_abs":"http://arxiv.org/abs/1804.11237v2","url_pdf":"http://arxiv.org/pdf/1804.11237v2.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":"deep-learning-improved-by-biological","repo_url":"https://github.com/takyamamoto/BRU_chainer","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":null}],"tasks":[{"task_slug":"deep-learning","task_name":"Deep Learning"}],"methods":[{"method_slug":"elu","method_name":"ELU"},{"method_slug":"relu","method_name":"ReLU"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}