{"title": "Effective Learning Requires Neuronal Remodeling of Hebbian Synapses", "book": "Advances in Neural Information Processing Systems", "page_first": 96, "page_last": 102, "abstract": null, "full_text": "Effective Learning Requires  Neuronal \n\nRemodeling of Hebbian Synapses \n\nGal Chechik \n\nIsaac  Meilijson  Eytan Ruppin \n\nSchool of Mathematical Sciences \n\nTel-Aviv University Tel  Aviv,  Israel \n\nggal@math.tau.ac.il \n\nisaco@math.tau.ac.il \n\nruppin@math.tau.ac.il \n\nAbstract \n\nThis paper revisits the classical neuroscience paradigm of Hebbian \nlearning.  We  find  that  a  necessary  requirement  for  effective  as(cid:173)\nsociative  memory  learning  is  that  the  efficacies  of  the  incoming \nsynapses  should  be  uncorrelated.  This  requirement  is  difficult  to \nachieve in  a  robust  manner by  Hebbian synaptic learning,  since it \ndepends on network level information.  Effective learning can yet be \nobtained by a neuronal process that maintains a zero sum of the in(cid:173)\ncoming synaptic efficacies.  This normalization drastically improves \nthe  memory  capacity of associative networks,  from  an  essentially \nbounded capacity to one that linearly scales with the network's size. \nIt also enables the effective storage of patterns with heterogeneous \ncoding levels in a single network.  Such neuronal normalization can \nbe successfully carried out by activity-dependent homeostasis of the \nneuron's synaptic efficacies, which was recently observed in cortical \ntissue.  Thus,  our findings  strongly suggest  that  effective  associa(cid:173)\ntive learning with Hebbian synapses alone is  biologically implausi(cid:173)\nble and that Hebbian synapses must be continuously remodeled by \nneuronally-driven regulatory processes in  the brain. \n\nIntroduction \n\n1 \nSynapse-specific changes in synaptic efficacies, carried out by long-term potentiation \n(LTP)  and depression  (LTD)  are thought to underlie cortical self-organization and \nlearning  in  the brain.  In  accordance  with  the  Hebbian  paradigm,  LTP  and  LTD \nmodify synaptic efficacies as a function of the firing of pre and post synaptic neurons. \nThis paper revisits the Hebbian paradigm showing that synaptic learning alone \ncannot  provide  effective  associative  learning  in  a  biologically  plausible \nmanner,  and  must  be  complemented  with  neuronally-driven  synaptic \nremodeling. \n\nThe  importance  of  neuronally  driven  normalization  processes  has  already  been \ndemonstrated in the context of self-organization of cortical maps  [1,  2]  and in  con(cid:173)\ntinuous  unsupervised  learning as  in  principal-component-analysis networks  [3].  In \nthese scenarios normalization is necessary to prevent the excessive growth of synap-\n\n\f", "award": [], "sourceid": 1774, "authors": [{"given_name": "Gal", "family_name": "Chechik", "institution": null}, {"given_name": "Isaac", "family_name": "Meilijson", "institution": null}, {"given_name": "Eytan", "family_name": "Ruppin", "institution": null}]}