Folding Forum

Jimbo is right. It's unlikely that FAH will use half=precision (16-bit) calculations in any way. Most of the existing calculations use standard Single Precision (32-bit). A limited number require Double Precision (64-bit).

The AI Cores are fine for 3D images to the accuracy that looks good a video screen and fine for estimating logical conclusions from an AI Training set -- or even for driving a self-navigating car -- but not for scientific calculations.

Let us get back to the topic.
Google/Deepmind just won the CASP competition using AI.
That proves that AI can know what is more plausible and thus guide F@H WU generation.

Deepmind already mentioned here viewtopic.php?f=15&t=31183

Maybe the already folded proteins by FAH could be used as input to AI to learn how it worked in the past. Then AI predicts the future for not folded proteins by FAH. If the AI predicted protein folds match what FAH outputs for these proteins then AI is good enough. If it is also faster than FAH then AI wins.

FAH and CASP have different goals. You are at a FAH support site not a CASP support site.
AI can help CASP but NOT FAH.

bruce wrote:AI can help CASP but NOT FAH.

You are assuming a fact not in evidence. Foldy said nothing about CASP. He was merely suggesting a way that AI might be used with FAH. I doubt that anyone has proven he is wrong yet.

"Yes, the CASP is aimed at predicting the normal structure. It's not surprising that AI would be good at that."
"AI can help CASP but NOT FAH."

Can you give reason for that assessment/opinion? I fail to see the fundamental difference.
Did you read about how they did it? Just because they did utilize AI in a way that does not apply to F@H does not say anything about the general applicability of AI or even the methodology of deepmind.
I agree with JimF, you have not made an argument for your case. I wrote multiple posts on how it could work for F@H.

@foldy:
The way you suggest does indeed not work because of the point bruce is making, but I did sketch how it could support the workings of F@H.

Deepmind used multiple NN one of which
"to predict a separate distribution of distances between every pair of residues"
That might save a lot of 'wiggling', would that prediction be used to bias the WU generation. Instead of following promising outliers one could generate likely hypothetical Markov states and look for transitions to known states.

I guess we need a scientist to answer that.

Interesting,
I don't know how accessible the AI is on Snapdragon chipsets.
It's usually a part sealed off within the CPU doing all kinds of low level background activity, like listening to whatever is spoken, convert the words to text, compresses it, and uploads it to ad and govt servers, for you to see ads about things you've just talked about,
Focusing camera, learning your life habit (where you go, what you do on a day to day basis, based on your location).
I hate AI, and as far as I know, a lot of the tracking feats are turned off on the Google Pixel 2 phones, as well as on the Essential phone,
But a lot of other phones like Samsung, etc, still have that stuff running in the background draining battery.

At least, that's the AI part of an ARM powered, modern android smartphone.
Not sure if there are other AI systems in desktop computer chips.

I used terms that I would like to link to those used in the Wikipedia article.
"native state" we called it normal structure.
"configuration space", "phase space" what I called state space.
"conformational states" we called them: states, current state, semi-stable state, quasi stable state, configuration cluster, attractor, local minimum
and I believe they are the same as Markov States.
"adaptive sampling" the part of F@H that happens on the server side generating the WUs.
That is the part that would use AI, to reduce the Samples = Trajectory = WU.

JimboPalmer Duckers ProDigit

You are way off topic, even in the wrong section of the forum. This thread it not about HW.

foldy wrote:Maybe the already folded proteins by FAH could be used as input to AI to learn how it worked in the past. Then AI predicts the future for not folded proteins by FAH. If the AI predicted protein folds match what FAH outputs for these proteins then AI is good enough. If it is also faster than FAH then AI wins.

You either did not read what bruce and I wrote before posting to this thread or did not understand. You also did not understand what F@H is about.

Please consider this recent publication https://www.technologyreview.com/s/6125 ... ges-in-ai/
FAH is attempting to model continuous processes so they can be better understood. AI (Deep Learning) is great at classifying objects where there are multiple examples that can be used for supervised learning.