[FRIAM] Can current AI beat humans at doing science?

Wed Jul 21 09:53:33 EDT 2021

I tend to side with Dave and Patrick on this issue, I guess. The following article provides a bit of an on-ramp to my perspective:

Beware explanations from AI in health care
https://science.sciencemag.org/content/373/6552/284

Their distinction is valid and sound. So, the authors would object to the idea that AlphaFold holds any substantive semantics. (By "semantics", I don't mean the modern computer science sense of that, but the more general philosophical sense.) But, in the spirit of true factoids, false narrative, I disagree with the gist of the article.

We can't explain the behavior of animals any better than we can explain the behavior of AlphaFold. In fact, we're much better at explaining the behavior of AlphaFold because, as Feynman suggests, we can *build* AlphaFold (... or some of us can, anyway), whereas we're still having some trouble with synthetic organisms.

But the fact that we can't understand animal behavior does NOT mean much. We've been NOT understanding animal behavior for millennia. Yet we make progress. So, in the end, I side with strong AGI position *and* reject the idea that AlphaFold is categorically different from prior technologies.

The resolution of that apparent contradiction is that I believe *all* explanation is via mimic models. Interpretable algorithms are a convenient fiction. (I.e. all models are always wrong.)

On 7/21/21 4:57 AM, Pieter Steenekamp wrote:
> Prof Dave West,
> 
> There is something different happening with the current generation of AI compared to the previous generation. The AI generation of Alan Newell @ Herb Simon, and I also want to include Big Blue that beat Gary Kasparov at chess was just encoding human intelligence and making it much faster. The AI did not contribute any novel concepts to the algorithm, the humans did that.
> 
> The current neural network based AI does add novelty to the solution. It learns and gains insight from the data in ways that humans can not.
> 
> Take AlphaFold for example, humans do not understand how to predict the folding of a protein by analyzing the amino acid sequence; it's beyond human understanding to do that. It's like AlphaFold looked at the 100000 odd known examples of amino acid sequences and the resulting folded protein structure and say - it's easy, you just look at this, then that, and the resulting folded protein is this. Even comprehending what AlphFold is saying is beyond human understanding - it's there to look at, it's in the weights of the gain and biases of the many connections of the artificial neural network, but humans can just not interpret it.
> 
> For a human to understand AlphaFold's reasoning to solve the protein folding problem is like expecting a 2 year old child to understand quantum mechanics. Or like me to understand my wife's mind.
> 
> AI does not have general intelligence, maybe it will never happen. But I think it's safe to say that in some narrow fields, like in the protein folding problem, AI is certainly more intelligent than humans. The important issue is that there is evidence that AI does add novelty to the solution.
> 
> Pieter
> 
>  
> 
> On Tue, 20 Jul 2021 at 22:46, Prof David West <profwest at fastmail.fm <mailto:profwest at fastmail.fm>> wrote:
> 
>     __
>     Thirty something years ago, Alan Newell walked into his classroom and announced, "over Christmas break, Herb Simon and I created an artificial intelligence." He was referring to the program Bacon, which fed with the same dataset as the human deduced the same set of "laws." It even deduced a couple of minor ones that Bacon missed (or, at least, did not publish).
> 
>     Simon and Newell tried to publish a paper with Bacon as author, but were rejected.
> 
>     AlphaFold (which I think is based on a program Google announced but has yet to publish in a "proper" journal) is, to me, akin to Bacon, in that it is not "doing science," but is merely a tool that resolves a very specific scientific problem and the use of that tool will facilitate humans who actually do the science.
> 
>     I will change my mind when the journals of record publish a paper authored by AlphaFold (or kin) as author and that paper at least posits a credible theory or partial theory that transcends "here is the fold of the xyz sequence to address why that fold is 'necessary' or 'useful'.
> 
>     davew
> 
> 
>     On Tue, Jul 20, 2021, at 1:12 PM, Pieter Steenekamp wrote:
>>     A year or so ago, Deepmind's AlphGo defeated the then world Go-champion Lee Sedol at a time when leading Ai researchers predicted it will be at least 10 years before AI can reach that level. But the valid question then was - why so excited? It's just a game. There is an interesting documentary on youtube about this at https://www.youtube.com/watch?v=WXuK6gekU1Y <https://www.youtube.com/watch?v=WXuK6gekU1Y>
>>
>>     What's happening now is that AI makes scientific discoveries beyond human ability. 
>>
>>     Is anybody worried where it will end?
>>
>>     I quote from https://www.nature.com/articles/s41586-021-03819-2 <https://www.nature.com/articles/s41586-021-03819-2>
>>     Highly accurate protein structure prediction with AlphaFold
>>     Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous experimental effort1–4, the structures of around 100,000 unique proteins have been determined5, but this represents a small fraction of the billions of known protein sequences6,7. Structural coverage is bottlenecked by the months to years of painstaking effort required to determine a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the 3-D structure that a protein will adopt based solely on its amino acid sequence, the structure prediction component of the ‘protein folding problem’8, has been an important open research problem for more than 50 years9. Despite recent progress10–14, existing methods fall far short of atomic accuracy, especially when no homologous structure is available. Here we provide the first
>>     computational method that can regularly predict protein structures with atomic accuracy even where no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14)15, demonstrating accuracy competitive with experiment in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates physical and biological knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm.
>>
>>
>>
>>
>>     - .... . -..-. . -. -.. -..-. .. ... -..-. .... . .-. .
>>     FRIAM Applied Complexity Group listserv
>>     Zoom Fridays 9:30a-12p Mtn GMT-6  bit.ly/virtualfriam <http://bit.ly/virtualfriam>
>>     un/subscribe http://redfish.com/mailman/listinfo/friam_redfish.com <http://redfish.com/mailman/listinfo/friam_redfish.com>
>>     FRIAM-COMIC http://friam-comic.blogspot.com/ <http://friam-comic.blogspot.com/>
>>     archives: http://friam.471366.n2.nabble.com/ <http://friam.471366.n2.nabble.com/>
>>
> 
>     - .... . -..-. . -. -.. -..-. .. ... -..-. .... . .-. .
>     FRIAM Applied Complexity Group listserv
>     Zoom Fridays 9:30a-12p Mtn GMT-6  bit.ly/virtualfriam <http://bit.ly/virtualfriam>
>     un/subscribe http://redfish.com/mailman/listinfo/friam_redfish.com <http://redfish.com/mailman/listinfo/friam_redfish.com>
>     FRIAM-COMIC http://friam-comic.blogspot.com/ <http://friam-comic.blogspot.com/>
>     archives: http://friam.471366.n2.nabble.com/ <http://friam.471366.n2.nabble.com/>
> 
> 
> - .... . -..-. . -. -.. -..-. .. ... -..-. .... . .-. .
> FRIAM Applied Complexity Group listserv
> Zoom Fridays 9:30a-12p Mtn GMT-6  bit.ly/virtualfriam
> un/subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
> FRIAM-COMIC http://friam-comic.blogspot.com/
> archives: http://friam.471366.n2.nabble.com/
> 

-- 
☤>$ uǝlƃ