[Lisa_teatalk] [Lisa_labo] Talk by Luca Rigazio on Wednesday

Guillaume Alain guillaume.alain.umontreal at gmail.com
Wed Mar 4 13:19:21 EST 2015


Hi everyone !

This is a reminder that we have a tea talk this afternoon with Luca Rigazio as guest, from Panasonic Research.

There seems to be some confusion about the actual time of the presentation, so we'll resolve the issue by having the presentation at 14:00. 

I'll do the usual sweep across the lab to remind everyone 5 minutes before we start. 

> On Mon, Mar 2, 2015 at 6:15 AM, Kyung Hyun Cho <cho.k.hyun at gmail.com> wrote:
> Dear all,
> 
> We have a talk this Wednesday by Luca Rigazio on convolutional neural networks at 14.00. See the details below.
> 
> Hope to see many of you there!
> - K
> 
> ---
> Luca Rigazio - Panasonic Silicon Valley Laboratory
> 
> ML background (MS, no Ph.D., sorry), I worked extensively in speech recognition/NLU, now I run PSVL in Cupertino, California, Panasonic's elite corporate innovation lab. Last year I caught the deep-learning virus (like many) and started working on it (like most), spun off a new dedicated team. We are still learning but enjoy doing cool stuff, I enjoy coding new algorithms with the team (mostly C++, don't like Matlab, starting to cope with Python).
> Today I will tell you a couple of things about us and our company, present some initial results especially in Vision, and hope to learn a lot from your great group!
> 
> --
> 
> Deep Clustered Convolutional Kernels
> 
> DNN's have recently achieved state of the art performance: in practice the network architecture has
> to be manually set by domain experts, generally by a costly trial and error procedure, which often
> accounts for a large portion of the final system performance.
> We view this as a limitation and propose a training algorithm that automatically optimizes network architecture, by progressively
> increasing model complexity and then eliminating model redundancy by selectively removing parameters at training time.
> For convolutional neural networks, our method relies on iterative split/merge clustering of convolutional kernels interleaved by stochastic gradient descent.
> We present a training algorithm and experimental results on three different vision tasks, showing improved performance compared to similarly sized hand-crafted architectures.
> 
> From my mobile
> 
> 
> _______________________________________________
> Lisa_labo mailing list
> Lisa_labo at iro.umontreal.ca
> https://webmail.iro.umontreal.ca/mailman/listinfo/lisa_labo

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://webmail.iro.umontreal.ca/pipermail/lisa_teatalk/attachments/20150304/c389b9ac/attachment.html 


More information about the Lisa_teatalk mailing list