[AndyGoode]

A few weeks ago I decided I should learn something about the popular new gene editing technique called CRISPR/cas9, so I watched about three videos from YouTube on this topic that were pretty clear...

()
What is CRISPR?

Bozeman Science
Published on Feb 18, 2016
()
How CRISPR lets you edit DNA - Andrea M. Henle

TED-Ed
Published on Jan 24, 2019
()
CRISPR Explained

Mayo Clinic
Published on Jul 24, 2018

Upon watching that first video, it hit me: "Hey! This is how nature actually does machine learning!" Admittedly the learning being done is not in a brain, but rather in a bacterial cell, and what is being learned is not anything visual, but rather chemical in the form of DNA strands, but still, since it's likely that natural learning systems started simply like this, it's possible the brain's learning mechanism that arose later in evolution uses the bacterial method as a foundation and just added a few modifications. I think that's a pretty hot idea to investigate if anyone's looking for an interesting research project for an article or a science fair project, especially since the way the brain learns at a higher level of abstraction is unknown, and is one of the most important questions in AI. I don't have time to work on this idea, but if that idea catches anyone's fancy and they decide to work on it, I'll chip in with any advice or ideas I have along the way as they develop it.

Here's the CRISPR learning process as I understand it so far, described chronologically:
1. All input with any DNA pattern is assumed to be hostile. Any input to the system has its pattern matched against other patterns the system has seen, which are stored in its CRISPR.
2a. If a match is detected, the input is destroyed before it can affect anything in the cell.
2b. Else if a match is *not* detected, a new "class" is created, and the new input represents that class, and a copy of that new input DNA is permanently stored in the CRISPR, but in a safe way, for match against future patterns. The input is then destroyed, as before. In either case, the input is destroyed under the assumption that it is hostile.

How might this relate to machine learning in a computer?
() Instead of the input consisting of DNA patterns (which are strings of the 4 symbols A, C, G, T) it might consist of strings of bits (which are the 2 symbols 0, 1).
() The matching algorithm would be almost identical to existing string matching functions used in C and other computer languages.
() Instead of matching a string verbatim, possibly only short samples of the string could be used as a method of generalization or of associative memory. This would be the tricky part, since verbatim string matching isn't particularly interesting, but generalization is, so it would require some ideas for what and how much to store--but on the other hand verbatim matching might still be interesting if the stored string were a single feature of the input rather than the entire input. In the latter case, the interest would be in which and how many features would flag a match.

Anyway, that should be a good start to fire someone's imagination, if this sort of topic interests them. Then maybe we all could see some more active projects going on in this forum.

[LOCKSUIT]

Thanks for sharing.

I read all above but it seems no different than my AGI plan. This seems to be simply pattern matching, generalization, memory of different properties. So, nothing new for me. However it did make me realize to slightly adjust something in my work/reanalyze, like a rule goes for all not just x, though I already knew about that really.

[AndyGoode]

This seems to be simply pattern matching, generalization, memory of different properties.

That's a very good point I hadn't considered: There are really no data dynamics involved in this system. Other than the low-level physical mechanics of moving RNA strands close to each other to see if they match, no data moves any more than it does in a neural network. In contrast, imagine mentally setting a block on top of another block: data of some sort must be moved to do that intelligent operation. Maybe the only exciting thing about my analogy is that for the first time we can see *one* solution to how nature solved the machine learning problem, right down to the exact molecular arrangement of the memory and incoming data. It's still possible, though, that the "few modifications" I mentioned that could be done to this system might be only the addition of dynamic data, whereupon the system might transcend one of the limitations of neural nets.

[LOCKSUIT]

I agree the 2 molecule strand pairs are comparing and joining as one because they match enough and are related from their commonalities being close enough. This pattern matching similarity detection is indeed the same in the brain, just these DNA strands do it as actual molecular strands, and indeed in the brain it looks just like that when a cell in duplicating and the pairs line up and split into 2 with chains like bridges between them, if you get me. These bridges are the commonalities/relations. It looks like 2 nodes with spider webs circle bursts magnetizing to each other like arms do to hug, and they pull each other in using those bridges/commonalities. Sorta looks like a fork end up against a fork end.

Good idea to try or simulate them molecules doing such to make some sort of relational brain naturally or some sort of new DNA/cells and processes.

[LOCKSUIT]

Also see this Two Minute Paper video too, it talks about GPT-2, and computation vs handcrafted:



My notes from it:
- brute force, searchdown, learning/training, efficient heuristic?.....vs.... handcrafted rule net white-box
- Always try to balance and add more compute than human knowledge structure in your algorithm.
- Try imagining if we had 100 times more compute, which are the most important parts in our algorithm that would benefit.

[AndyGoode]

Thanks for the video, Locksuit. I definitely agree with that video's conclusions when working on applied AI on a digital computer, but I think we should keep in mind that the brain is a prediction machine, as Jeff Hawkins called it, so processing speed may not be as important as memory. This would be true of AGI in general, in contrast to applied AI.

(p. 89)
   "Prediction" means that the neurons involved in sensing
your door become active in advance of them actually receiving
sensory input. When the sensory input does arrive, it is com-
pared with what was expected. As you approach the door, your
cortex is forming a slew of predictions based on past experi-
ence. As you reach out, it predicts what you will feel on your
fingers, when you will feel the door, and at what angle your
joints will be when you actually touch the door.

(p. 89)
We are making continuous
low-level predictions in parallel across all our senses.
   But that's not all. I am arguing a much stronger proposition.
Prediction is not just one of the things your brain does. It is the
primary function of the neocortex, and the foundation of intelli-
gence. The cortex is an organ of prediction. If we want to under-
stand what intelligence is, what creativity is, how your brain works,
and how to build intelligent machines, we must understand the
nature of these predictions and how the cortex makes them.
Even behavior is best understood as a by-product of prediction.

Hawkins, Jeff. 2004. On Intelligence. New York: Times Books.

[LOCKSUIT]

I agree, predictions find shortcuts that we 'expect'. Is actually faster than our current Compute.