{"pk":31719,"title":"Principal Hidden Unit Analysis: Generation and Interpretation of Principal Networks by Zminimum Entropy Method","subtitle":null,"abstract":"In the present paper, a principal hidden\nunit analysis with entropy minimization is\nproposed to obtain a simple or fundamental\nstructure from original complex structures.\nThe principal hidden unit analysis is com-\nposed of four steps. First, entropy, defined\nwith respect to the hidden unit activity, is\nminimized. Second, several principal hidden\nunits are selected, according to il-index, rep-\nresenting the strength of the response of hid-\nden units to input patterns. Third, the per-\nformance of the obtained principal network is\nexamined with respect to the error or generalization. Finally, the internal representa-\ntion of the obtained principal network must\nappropriately be interpreted. Applied to a\nrule-plus-exception, a symmetry problem and\nan autoencoder, it was confirmed in all cases\nthat by using entropy method, a small num-\nber of principal hidden units were selected.\nWith these principal hidden units, principal\nnetworks were constructed, producing targets\nalmost perfectly. The internal representation\ncould easily be interpreted especially for simple problems","language":"eng","license":{"name":"","short_name":"","text":null,"url":""},"keywords":[],"section":"Submitted Presentations","is_remote":true,"remote_url":"https://escholarship.org/uc/item/9c0002vr","frozenauthors":[{"first_name":"Ryotaro","middle_name":"","last_name":"Kamimura","name_suffix":"","institution":"Tokai University","department":""}],"date_submitted":null,"date_accepted":null,"date_published":"1993-01-01T18:00:00Z","render_galley":null,"galleys":[{"label":"PDF","type":"pdf","path":"https://journalpub.escholarship.org/cognitivesciencesociety/article/31719/galley/22787/download/"}]}