CoMeT | <%=title %>
Collaborative Management of Talks Hello! sign in or register
Bookmark Talks, Share with Friends, and We Recommend More!
Advanced Search
Talk Detail
Posted: ISP Admin on  Nov 26 01:20:15 PM
Title: NETLAND: Visualization of THE GEOMETRY AND DYNAMICS OF Hidden Unit Space  
Paul Munro
School of Information Sciences, University of Pittsburgh
Sponsor: University of Pittsburgh  >  School of Computing and Information  >  Intelligent Systems Program
Series: ISP Colloquium Series
Date: Dec 13, 2019 12:30 PM - 1:00 PM
Location: 5317 Sennott Square
Groups Posted: Intelligent Systems Program Machine Learning Group 
Bookmarked by:  PatHealy peterb

Classification by multi layer neural networks depends on the existence of appropriate features in the early hidden layers, so that the representations are linearly separable in the penultimate layer.   By using hidden layers with just two or three units, the representational structure of the intermediate layers can be visualized.  Time courses of the evolution of the hidden layer representations are visualized as animations.  The visualizations reveal tendency for the hidden unit image of the input space to collapse into a non-linear (warped) manifold of lower dimensionality; i.e. the weight matrix is (nearly) singular.  A task which is not linearly separable in the input space is rendered linearly separable by the warping of the manifold.  Because of the matrix singularity, deeper layers of the network do not have sufficient information to reconstruct the input in its original form.  Thus, the deep layers are incapable of discriminating certain distinct stimulus patterns.

People Who Bookmarked This Talk, Also Bookmarked
RSS Feed: RSS 2.0
ATOM Feed: Atom
iCalendar: iCal
Share: Bookmark and Share
Google Calendar:
CoMeT Blog
©2009-2020 CoMeT - Supported by Google Grant
School of Information Sciences, University of Pittsburgh, 135 North Bellefield Avenue, Pittsburgh, PA 15260