Page 19 of 30

Re: If you are bored with Deep Networks

Posted: September 4th, 2018, 9:39 am
by ISayMoo
Do the Wolfe conditions apply to a stochastic optimisation setting?

Re: If you are bored with Deep Networks

Posted: September 4th, 2018, 12:35 pm
by Cuchulainn
Do the Wolfe conditions apply to a stochastic optimisation setting?
Good question to get onto radar. A guess is 

probabilistic WC
stochastic WC

wouldn't surprise me one bits, so it wouldn't :)

Re: If you are bored with Deep Networks

Posted: September 4th, 2018, 4:18 pm
by ISayMoo
Would weak Wolfe conditions be enough?

Re: If you are bored with Deep Networks

Posted: September 4th, 2018, 5:25 pm
by ISayMoo
OK, so this looks like a reasonably decent paper: Probabilistic Line Searches for Stochastic Optimization

They discuss convergence guarantees briefly in Sec. 3.4.1. Experiments look encouraging, but I'd like to test them on something more challenging.

EDIT: replaced the URL with a published version.

Re: If you are bored with Deep Networks

Posted: September 5th, 2018, 12:44 am
by katastrofa
Ugh, can't you simply use a quantum annealer?

Re: If you are bored with Deep Networks

Posted: September 5th, 2018, 7:24 am
by ISayMoo
I can't afford one.

Re: If you are bored with Deep Networks

Posted: September 10th, 2018, 8:49 pm
by berndL
Over 10% of it is dog photos, and less than half of that is cats. While everyone knows that most of Internet is cat pics!
Great statistics. I didnt knew this. But seems people like cats more than dogs. Sorry  for beeing off topic though.

Re: If you are bored with Deep Networks

Posted: September 11th, 2018, 6:27 am
by Cuchulainn
There was a time when baboons' derrieres were the benchmark for CAD screen acceptance.


Image

Re: If you are bored with Deep Networks

Posted: September 11th, 2018, 9:10 am
by Cuchulainn
Seriously many of the benchmark and articles tend to focus on pixel-related and vision? Is this not stunting the growth of ML in other areas? Is it engendering tunnel vision?

The mathematical rigour is very patchy. Just look at activation functions and ask: what, what kinds of, why, how, why, when, relationships, what if etc.

This is a crucial component and so little is known about it.

Re: If you are bored with Deep Networks

Posted: September 11th, 2018, 11:29 am
by ISayMoo
Computer Vision is a very old topic with many practical applications, so it's natural that it receives a lot of interest. But there is a lot of research in other areas: speech processing, text processing, robot control.

Re: If you are bored with Deep Networks

Posted: September 11th, 2018, 11:38 am
by Cuchulainn
Computer Vision is a very old topic with many practical applications, so it's natural that it receives a lot of interest. But there is a lot of research in other areas: speech processing, text processing, robot control.
I don't doubt that. Speech processing is also ancient. Is there more? That was really my question, not a summary of current applications.
And whether those 60s/70s methodologies are suitable going forward.  How do you support geometry and topological data?

I read about HearSay-II during s/w designs

http://i.stanford.edu/pub/cstr/reports/ ... 6-1123.pdf

Re: If you are bored with Deep Networks

Posted: September 11th, 2018, 1:40 pm
by ISayMoo
Modern speech processing is not 60s/70s anymore ;-) Neither is vision.

Re: If you are bored with Deep Networks

Posted: September 11th, 2018, 2:45 pm
by Cuchulainn
Modern speech processing is not 60s/70s anymore ;-) Neither is vision.
Of course. The difference is the documentation.

Re: If you are bored with Deep Networks

Posted: September 11th, 2018, 5:14 pm
by ISayMoo
Not really. I'm getting slightly tired of this ping-pong to be honest ;-)

Re: If you are bored with Deep Networks

Posted: September 11th, 2018, 5:42 pm
by Cuchulainn
Not really. I'm getting slightly tired of this ping-pong to be honest ;-)
Mostly pong to date, I'm waiting on some ping.