Coast-to-Coast Seminars

  • Posted October 20, 2009. Last modified August 4, 2010 by

Today I found out about some cross-Canada computer / math seminars that are sponsored by a consortium of research / computing organizations across Canada including SHARCNET, WestGrid, ACENet, IRMACS and AARMS. Today’s talk was given at the University of Toronto by Prof. Geoffrey Hinton (see the SHARCNET site for a description / abstract).

The first part of the talk showed a network learning how to recognize written characters. The approach was unique (at least to what I’ve encountered in my soft computing class) in that it didn’t assign and back-propagate the labels from the start. The network learned “features” based on the input patterns and then assigned the labels after the pattern had been learned. The coolest feature was being able to “visualize what the network is thinking” by doing the process sort of in reverse. The second part of the talk applied a similar technique to the motion of person wearing sensors. The network could be trained to recognize the style of motion of the person and then from that, new styles of walking could be “imagined” by the network. For example the network could image the person changing walking styles midwalk through the visualization even though it hadn’t been trained in this way. In all it was very interesting. It would be fun to try to apply some of these techniques to wireless networks. Perhaps the motion modeling could be applied to mobile wireless devices to help with hand-offs?

Anyway, if anyone is interested and you are at one of the Universities which is a part of the groups putting these on, they happen every other week. You can see at schedule at the SHARCNET website, or probably at the group you are a part of at your school. As far as I know, anyone can attend!

One comment on “Coast-to-Coast Seminars”

  1.  

    About the character recognition… Clustering before classification. That's a good idea… I'll put a sticky up in my brain about that. Interesting– this must be a form of compression, wherein we have reduced the dimensionality of the problem by removing information from the characters to recognize. I'll have to read that abstract 😀

Leave a Reply

Your email address will not be published. Required fields are marked *