BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Journal Club: Visualizing Data using t-SNE - Emli-Mari Nel (Univer
 sity of Cambridge)
DTSTART:20091127T111500Z
DTEND:20091127T120000Z
UID:TALK21317@talks.cam.ac.uk
CONTACT:Emli-Mari Nel
DESCRIPTION:\nIn this week’s journal club I would like to discuss the fo
 llowing paper:\n\n"Visualizing Data using t-SNE" by Laurens van der Maaten
  and Geoffrey Hinton\,\nJournal of Machine Learning Research\, Vol. 9 (Nov
 ember 2008)\, pp. 2579-2605.\n\nwww.cs.utoronto.ca/~hinton/absps/tsnefinal
 .pdf\n\n\n"We present a new technique called “t-SNE” that visualizes h
 igh-dimensional data by giving each\ndatapoint a location in a two or thre
 e-dimensional map. The technique is a variation of Stochastic\nNeighbor Em
 bedding (Hinton and Roweis\, 2002) that is much easier to optimize\, and p
 roduces\nsignificantly better visualizations by reducing the tendency to c
 rowd points together in the center\nof the map. t-SNE is better than exist
 ing techniques at creating a single map that reveals structure\nat many di
 fferent scales. This is particularly important for high-dimensional data t
 hat lie on several\ndifferent\, but related\, low-dimensional manifolds\, 
 such as images of objects from multiple classes\nseen from multiple viewpo
 ints. For visualizing the structure of very large data sets\, we show how\
 nt-SNE can use random walks on neighborhood graphs to allow the implicit s
 tructure of all of the\ndata to influence the way in which a subset of the
  data is displayed. We illustrate the performance of\nt-SNE on a wide vari
 ety of data sets and compare it with many other non-parametric visualizati
 on\ntechniques\, including Sammon mapping\, Isomap\, and Locally Linear Em
 bedding. The visualizations produced by t-SNE are significantly better tha
 n those produced by the other techniques on almost all of the data sets."\
 n
LOCATION:TCM Seminar Room\, Cavendish Laboratory\, Department of Physics
END:VEVENT
END:VCALENDAR