I saw one of the papers posted on the news section of the site, discussing the potential for deep learning in drug discovery: http://folding.stanford.edu/2017/07/10/ ... or-pharma/
I was a little surprised by this because I thought that F@H basically did statistical modeling, not machine learning or deep learning. So now I'm wondering, do any of the cores do ML work? Might they in the future?
Does FaH do any ML / deep learning computations?
Moderators: Site Moderators, FAHC Science Team
-
ChristianVirtual
- Posts: 1576
- Joined: Tue May 28, 2013 12:14 pm
- Location: Tokyo
Re: Does FaH do any ML / deep learning computations?
Right now I would assume no available core is doing ML; they are “just” tools to compute the proteins and ML might happen before and after a WU to steer the initial setup of projects and to analyze the results.
But who knows what is cooked behind the curtain.
In some other forum there was recently some soliciting for a BOINC/FAH style of new AI community : but seems not yet took off. Funny; they asked in multiple distributed computing forums for unused cycles ; made me smile. What is an unused cycle ? Don’t have any.
But who knows what is cooked behind the curtain.
In some other forum there was recently some soliciting for a BOINC/FAH style of new AI community : but seems not yet took off. Funny; they asked in multiple distributed computing forums for unused cycles ; made me smile. What is an unused cycle ? Don’t have any.
Re: Does FaH do any ML / deep learning computations?
Haha that is funny.
-
JimboPalmer
- Posts: 2522
- Joined: Mon Feb 16, 2009 4:12 am
- Location: Greenwood MS USA
Re: Does FaH do any ML / deep learning computations?
https://arstechnica.com/civis/viewtopic ... &t=1400871
You will notice he has no goal other than building the largest. He is not planning on solving any of mankind's ills, he just wants to stroke his ego.
I was not kind to to him.
You will notice he has no goal other than building the largest. He is not planning on solving any of mankind's ills, he just wants to stroke his ego.
I was not kind to to him.
Tsar of all the Rushers
I tried to remain childlike, all I achieved was childish.
A friend to those who want no friends
I tried to remain childlike, all I achieved was childish.
A friend to those who want no friends
Re: Does FaH do any ML / deep learning computations?
A traditional computer can do some things very quickly and efficiently and is very slow and inefficient at other things. One important difference is whether the goals/questions are highly focus or highly unfocused or somewhere in between. The brain tends to prefer the opposite.
(Read about the Google Cat project. Use 16,000 computers for a month with the goal of "Find the cat in pictures")
If the goal is to train a self-driving car to avoid killing pedestrians, we could turn a lot of them loose on the road and whenever a pedestrian was killed, we's (symbolically) say "Bad Dog" but there are better ways, even though that same technique (maybe using "Good Dog") can teach a puppy to fetch a ball.
FAH doesn't use pure guided/focused searches. It uses a semi-guided technique. The typical project starts with a random sampling of a protein's motions. Then using a Bayesian process, interesting areas for study are separated from uninteresting areas. This makes the focused (sequential) processing much, much more effective. This refinement process may be repeated on a given protein.
FAH researchers have made many, many improvements in the 15+ years that FAH has been going. FAH's management are good about reading research papers or attending conferences about anything that might improve FAH's productivity, whether or not the latest buzz-words are including in the associated title.
(Read about the Google Cat project. Use 16,000 computers for a month with the goal of "Find the cat in pictures")
If the goal is to train a self-driving car to avoid killing pedestrians, we could turn a lot of them loose on the road and whenever a pedestrian was killed, we's (symbolically) say "Bad Dog" but there are better ways, even though that same technique (maybe using "Good Dog") can teach a puppy to fetch a ball.
FAH doesn't use pure guided/focused searches. It uses a semi-guided technique. The typical project starts with a random sampling of a protein's motions. Then using a Bayesian process, interesting areas for study are separated from uninteresting areas. This makes the focused (sequential) processing much, much more effective. This refinement process may be repeated on a given protein.
FAH researchers have made many, many improvements in the 15+ years that FAH has been going. FAH's management are good about reading research papers or attending conferences about anything that might improve FAH's productivity, whether or not the latest buzz-words are including in the associated title.