[4eyes] MS Project Adam

[Chris Sweeney]

Very interesting! You have to figure that Microsoft (and any other decently
large company) is investing heavily in deep learning so this shouldn't be
too much of a surprise. Claiming that Adam is one of the few systems that
can handle 22k images is a joke though... there are CVPR and NIPS papers
the past two years even that do this level of input on a single machine.

The HOGWILD! idea is interesting though. Training these massive models
involves a lot of random perturbation to avoid local minima, so allowing
machines to write over each other in memory is just another dimension of
randomness. It is interesting that it improves accuracy though. Hopefully
they'll release more details or publish papers describing some techniques


On Mon, Jul 14, 2014 at 3:37 PM, Matthew Turk <mturk at cs.ucsb.edu> wrote:

> Microsoft’s new Project Adam – deep learning and computer vision:
> “Brain-scaled neural networks”
>
>
>
> http://www.wired.com/2014/07/microsoft-adam/
>
>
>
> _______________________________________________
> Ilab-users mailing list
> Ilab-users at lists.cs.ucsb.edu
> https://lists.cs.ucsb.edu/mailman/listinfo/ilab-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cs.ucsb.edu/pipermail/ilab-users/attachments/20140714/6216838e/attachment.html>