Deep learning in recent VAE paper

Moderators: Site Moderators, PandeGroup

Deep learning in recent VAE paper

Postby wishtoremainanon » Sun Dec 30, 2018 10:53 pm

In article 192 here:
https://foldingathome.org/papers-results/

It notes is use of deep learning for compression.
In the case that this is added to the folding at home software,
and before there are some sort of public guidelines for deep learning
then it might be nice to have some sort of
A. note that 192 is included in the project
or
B. note how many neurons are being used
for those who might be slightly cautious about
putting some AI on their computer.
wishtoremainanon
 
Posts: 3
Joined: Fri Nov 09, 2018 11:20 pm

Re: Deep learning in recent VAE paper

Postby bruce » Mon Dec 31, 2018 2:43 am

This is a peer-reviewed research paper written by the scientists who extract data from results obtained by those wo use FAH software. The Donors who participate in FAH use various FAHCores to analyze how proteins fold. Those results are returned to the server where data reduction and compression takes place. Paper 192 describes the type of things that happen (or might happen in the future) on the back-end servers. The current version of the FAH software that is being distributed and used by the FAH Donors uses conventional compression techniques.
bruce
 
Posts: 21871
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Deep learning in recent VAE paper

Postby foldy » Mon Dec 31, 2018 12:09 pm

wishtoremainanon wrote: B. note how many neurons are being used
for those who might be slightly cautious about
putting some AI on their computer.

Are you affraid of AI? Current AI is only auto generated algorithms to filter and categorize some data. So if AI is trained to recognize cats on images and you show it another cat image it can recognize a cat again. Or if you train AI to compress some data and you give it such data then it can compress it again. It has nothing to do with real intelligence. It can only do what it is trained for like any other program on your PC.
foldy
 
Posts: 1246
Joined: Sat Dec 01, 2012 3:43 pm

Re: Deep learning in recent VAE paper

Postby ProDigit » Mon Dec 31, 2018 12:15 pm

I read graphics cars with deep learning technologies are not only insanely expensive, but in terms of Bitcoin mining, perform extremely poorly, or not at all.
I thought the same would be true for folding.

I thought deep learning, or ai, is about content creation, not replication like fah.
ProDigit
 
Posts: 194
Joined: Sun Dec 09, 2018 10:23 pm

Re: Deep learning in recent VAE paper

Postby wishtoremainanon » Mon Dec 31, 2018 9:58 pm

foldy wrote:
wishtoremainanon wrote: B. note how many neurons are being used
for those who might be slightly cautious about
putting some AI on their computer.

Are you affraid of AI? Current AI is only auto generated algorithms to filter and categorize some data. So if AI is trained to recognize cats on images and you show it another cat image it can recognize a cat again. Or if you train AI to compress some data and you give it such data then it can compress it again. It has nothing to do with real intelligence. It can only do what it is trained for like any other program on your PC.


Until further information, yes I am a little bit wary of knowing the complexity of AI I personally run.
Also, I am not sure how fast they turn the research into code, but I can
imagine that the process hypothetically could be fast.
wishtoremainanon
 
Posts: 3
Joined: Fri Nov 09, 2018 11:20 pm

Re: Deep learning in recent VAE paper

Postby foldy » Tue Jan 01, 2019 1:30 pm

You should only worry about AI driving a car or similar where real people can get hurt. The AI training even does not run inside the car but only at data centers with much compute power and big training data. Only the generated algorithm e.g. for detecting road signs then gets loaded into an AI car and cannot change there. But it is still only the next level of automation - these AI cannot do anything on its own but only what it is programmed for. It is the same for algorithms running at your PC.

If anything with AI could get wrong like it builds a conciousness and sends its terminators against humanity then this will happen in the big data centers where AI gets trained with millions of GPUs and data sets - and not at your PC.

But like bruce said even if they apply this algorithm to FAH it does not happen at your PC but at the Stanford University servers.
foldy
 
Posts: 1246
Joined: Sat Dec 01, 2012 3:43 pm

Re: Deep learning in recent VAE paper

Postby bruce » Tue Jan 01, 2019 8:27 pm

The problem with AI for driving is that it ONLY learns from experience. If you show it thousands of pictures of cats and dogs and tell it which are which, it will do a pretty good job of guessing the most probable answer to "Is it a cat or a dog?" If you show it an elephant, it will choose one of the two answers it knows.


A self-driving car sees many cases where other people follow the conventional rules of the road. It's very likely to do the wrong thing if it encounters something unexpected -- like a wrong-way-driver coming at you because it has never encountered that type of situation before. Then the developers have to add training examples to teach it what to do.
bruce
 
Posts: 21871
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.


Return to The Science of FAH -- questions/answers

Who is online

Users browsing this forum: No registered users and 1 guest

cron