Chris says it is controversial to say things like name of this topic.
Is it really something controversial?

What meditation might be other than biological version of machine learning?

Hi Ni Nurta

I guess machine learning might be a metaphor for conditioning the mind to move a certain way, 'see' a certain way.

If that is the case then maybe the metaphor makes sense to some people. But it doesnt make sense to me because I am not a computer scientist.

Perhaps Chris wanted you to expand on the topic so that folks who dont understand computer science, like me, could understand the metaphor. I dont live inside Chris' head so I can only speculate.


But I must say .. a lot of your metaphors/similes/analogies are beyond my understanding .. and thats absolutely ok by me

Ni Nurta --

​​​​​​​I said what you said was a bold claim. I'd like to hear you explain why you said this, beginning with a description of machine learning and then a description of how your comparison makes sense. I'm asking for this in the interest of coming to at least some level of understanding of what you meant. Unless we can reach that point, I'll have to retain my current opinion. Rigvht now, I don't see the similarities. This may be due to my short-sightedness, ignorance of one or the other terms you used, or some other misunderstanding about the meanings of those terms. I'd like to "get it" for once.

Assuming some metaphysics where the brain exist, meditation is changing the brain. I am not sure calling it machine learning gives you anything.

What does your definition of machine learning do that is different from normal learning or programming the brain?

I'm fascinated by this question, but also think the burden's on Ni Nurta to explain why meditation, in particular, is like machine learning!

Machine learning is already similar to how the brain works, simply because neural networks are explicitly modeled on a simplified interpretation of brain functioning. A massive network of "neurons" (whether nerve cells or software functions) accept an incoming signal of some sort. The individual neurons respond to the signal in a very simple way, often just a yes or no, or by modifying the signal in some way. In vast aggregate, in complex arrangements of interconnections, these simple responses add up to remarkably intelligent behavior. The neurons are simple; the networks are complex.

Learning happens because individual neurons remember values that they use in responding to incoming signals. These are usually called "weights" in machine learning. You train a model by putting in a signal, testing whether the outcome is correct or not, and continually adjust the weights of all the various neurons so that the outcome slowly hones in on the result you want. This is called training, and the rate at which it adjusts weights is called the "learning rate". Training can take months, and burn the electrical budget of a small country, because they are adjusting billions of parameters, and running millions of training iterations.

Once a model is trained, it is essentially "frozen": you don't update the weights again, you just put signal through, and expect that the answer is correct. When Google asks you to click on all the pictures with a mailbox in it, you are providing test data for training its models. It puts the pictures into its model, asks "mailbox?" and if the answer doesn't match what you told it, it updates the weights and tries again.

The human mind (we're drifting off into my own interpretations here) does the same thing, essentially, except that instead of having a clear distinction between the "training" stage and the "using it" stage, we kind of mush the two together. During the course of the day we're probably more "using it", maybe while dreaming we're more "training".

Google et al have a very clear criteria for training the success or failure of its model: is there a mailbox there or not? Did the Tesla see the stop sign or not? Human beings have a more complex set of criteria for training. Our neural networks take in data, we use that data to construct a model of our current context. We have positive or negative reactions to what we think we're experiencing, establish goals based on those reactions, and we form a course of action to reach those goals. We take action, and then evaluate the results of our action. Depending on the results, we adjust our weights: we become more or less likely to take a similar course of action in the future. I expect this is starting to sound familiar!

Training an ML model requires an army of costly GPU units because they're trying to zoom through a process that we human beings slowly churn through over the course of lifetimes.

I would love to know if Ni Nurta essentially agrees with the above; I'm hoping it can accelerate us past some of the basic questions. A more interesting question I would like to ask is -- given that the human mind could already be described as doing something very akin to machine learning, what is it about meditation in particular that is "more" like machine learning?

I would say it is the reverse: machine learning imitates what human learning appears like for humans. This leads to explanatory theories, based on phenomenal observation, which can take many metaphysical forms (it could be: "the process of meditation, like learning in general, is a gift of god"), but in this case seems to also involve neurophysicalist assumptions, ie, "the cause of these phenomena of human learning is brain mechanisms", which then leads to technological inventions that strive to imitate these theoretical mechanisms, in order to yield something that resembles the original phenomenon: human learning. The epistemological and ontological order in which this happens can then be conceiled or forgotten, and reversed when the metaphysical theories underpinning this practical process ("brain process is what's actually going on when experience is going on, and ML imitates brain thus meditatin is a variety of ML because physical mechanisms are more ontologically primary than experience") are reified into "true images of reality out there", to then yield propositions like the title of the present topic. This is a very common thing (i.e., mistake) that happens in science. Someone like Michel Bitbol, MD, quantum physicist, and phenomenologist who was part of the Mind&Life conferences by Varela, his then colleague at the Polytechnic school in Paris, explains this very well in his various papers, which you can find online. I recommend, for instance, "Towards a phenomenological constitution of quantum mechanics: a QBist approach," or "A merleaupontian ontology for quantum physics." Whether Ni Nurta is interested in actually engaging in open-minded discussion about this, though, or question his own assumptions, views and conclusions, is a much more ambiguous question in my view ;)

While I wait for Ni Nurta to educate me on his proposition...

The gap between physicalist science and consciousness has yet to be bridged. We have so little in the way of knowledge about what consciousness is and how it is produced that anything anyone proposes in regard to physical explanations is just speculation. This is possibly the modern paradigmatic example of an area for which the phrase "I don't know" is fully appropriate.

Ni asks "What meditation might be other than biological version of machine learning?" and some answers to the question would be that it might be a biological version of manual reweighting of a neural network, or a disk defragmenting operation, or a self-cleaning cycle, or a massage, or lots of things than other machine learning that can impact the functioning of a system. I'm not saying that it is any of these things, and I am not saying that it is not biological version of machine learning. I am just saying there are other possible answers, which means that a fun conversation could take place on this topic. 

While this is not an answer to Ni's question, at all, but just a tangentially connected thought, one thing that I find interesting about generative AI is that it might help people with no meditation experience understand (by seeing it done) how verbal thoughts can be produced without the need for a thinker. 

Also on the topic, some people might be interested in Doug Smith's talks on Buddhism and AI. He is not a computer scientist but rather a scholar of early Buddhism but he is clearly quite fascinated by this topic his analyses are fun and thought-provoking (which might be to say they serve as good prompts for the generative NI in my own head).

https://www.youtube.com/playlist?list=PL0akoU_OszRgY_O5tt-3mRipYsKWFREtt

I certainly hope it didn't look like I was trying to say that human learning imitates machine learning (how could that be?), or that machine learning represents some "closer to the metal" (pun intended) truth about the nature of reality. The first step of machine learning is finding a way of representing signal as grids of floating-point numbers, so obviously we are already many, many levels of representation deep. My post was just meant to give some background on the technological side of things. I have no opinion on the question of whether consciousness is ultimately mechanistic or deterministic.

Thanks for the pointers to both Doug Smith and Michel Bitbol! Any friend of Varela's is a friend of mine.

Hey,

As for me, I was merely adressing the OP

The question of whether there is any "thing" such as consciousness, or anything other than "consciousness", or both the same and different, or neither the same nor different, as consciousness, is interesting to ponder, i find.

Best,

O

https://www.youtube.com/watch?v=L_4BPjLBF4E

This reminds me of me trying to develop good habits/trying to meditate

Isnt learning the biological version of machine learning?

There is right a huge thing called computational psychiatry -- basically using large NN systems as proxies for human behavior, as I understand it. Turns out a lot ot features of NNs that are related to their size can be replicated in the human brain, and in fact are analogous. The example given was learning rate (in my mind -- sorry for the jargon -- learning rate is just update size parameter for an iterative process of successive approximation; it's not a concept that requires machine learning to exist, and is at least several centuries older). I think stressing the relation with ML is a bit misleading, though -- all those concepts ultimately come either from differential calculus or from information theory.

I think ultimately we will understand the mind as a very high-dimensional manifold with some sort of transport (like an affine connection, I suppose? https://en.m.wikipedia.org/wiki/Affine_connection ) defined on it, that gradually changes/evolves as the brain changes in time and connections form/are forgotten. At least this is my incompetent way of understanding the main idea of computational psychiatry. The link with ML is that there we are also moving along a very high-dimensional manifold (the "error surface"), trying to find a local minimum via gradient descent. I guess maybe if you consider meditation as being in some way extremely simple compared to normal function (which is a whole transport network), then it can be compared to a single run of gradient descent, on grounds of their simplicity? 

Okay maybe like this. In Chan thwre is a tradition of regarding all speech as error, and in fact all action as error. I've found in Seung Sahn's book a notion that I initially really liked, of actions being either good, bad or correct -- which is neither good nor bad. I'm not sure if I should even be allowed to think in these terms -- it seems like a very advanced teaching. 

Gradient descend minimizes error by findjng lowest point of a surface. In meditation we minimize the error of our minds by minimizing their superfluous activity. It is a certain kind of descent, and I have a feeling this notion could be made rigorous using the recent language of neural manifolds.

I'm thankful for opportunity to engage with these ideas on this forum.

​​​​​​​

Or.... we won't. Which is my incompetent way of re-emphasizing Seung Sahn's brilliant use of the phrase "Only don't know."

But we can, indeed, ponder.

First: Is there a need for consciousness? Does it stand to be explained? 

I think (but of course dont know for sure) what Mi Nurta was getting at was a certain dynamic conception of knowledge. From my experience, meditative or otherwise, all real knowledge exists *in order to do something*. You don't know something just because you're idly curious with it. What you obtain by idle curiosity is not knowledge, but a sort ot milquetoast fascination that is never able to penetrate into a thing's core. This is what third-rate professors show when they talk with their students about a subject that they don't really know, "oh its so complicated but we dont need it" wink wink nudge nudge cmon kids get down on my level. They have supposedly devoted their lives to studying the thing, and yet. 

This is also why people who don't have "skin in the game" generally perform worse on any game. Or why rich traders who "trade for fun" generally don't make much money. Etc. 

I think the point is that awakening is not an experience, but an action. Granting that, focusing on expwrience of being awakened will give you the experience of having an experience of being awakened. Focusing on being awakened will give you the experience of being awakened. You can differentiate between the two by asking, which one doesnt vanish when you look away from it? When you stop looking at a plant, your experience of it vanishes, but it itself does not.

Granting this dynamic conception of knowledge -- all knowledge comes from necessity -- you can see how a belief in consciousness might not be necessary for human beings. For me the final straw wrt that idea was when I looked at Pantone Color of the Year 2023, which was Viva Magenta. Look at Google Images: some of the examples are clearly a very deep red, and others a very beight violet. The colors are world apart, and corresponding light wavelenghts are as well. The issue is that despite the color spectrum being a 1D line, people actually see the color *wheel*, like a circle -- and they have magenta as this fake color in the place where the two ends are joined to make a circle. It literally exists only to keep you from being too confused and keep your user experience smooth. Its function is exactly and precisely the same as that of, say, windows in Microsoft Windows. IMO reducing mind to consciousness is like reducing computer science to UX -- done by most, but not advisable if you want to go to the deep end.

At least thats what I think about it. Please feel free to ignore my rambling! 

In my experience, meditation ultimately leads us to the simplest thing of all. It's not complicated, but the opposite. And I'm still interested in hearing from Ni Nurta - to explain what he meant, all by himself. 

Well, the point of eg computational psychiatry is not to be true. It needs to be true only insofar as it helps people. My original interest with it was because I was fascinated with treatments for epilepsy that come from it (right now, in July 2023, what they do if you have a tough case of epilepsy is they create a gigantic model of your brain, test various tough interventions on it, and do the one that worked the best on you. I mean thats crazy.), and because I was contemplating a career in medical applications of ML. Not because I wanted to know the truth. 

Truths can be varied - you wanted to pursue a pragmatic truth. A useful truth that allows lives to be enhanced. There's nothing wrong with that.

I think what you're both hitting at here, and something that needs to be kept front and center in discussions of ML versus the mind, is that all iterative improvements to a system are improvements towards achieving some particular goal. We are not zeroing in on some ultimate truth, we're just improving a tool's fitness for purpose. What those purposes are, and if and how they change, are an entirely separate question.

But when you start differentiating various kinds of truth, you're already doing it with a goal in mind. Probably to think more clearly about stuff. But why do that? To communicate better. That's what I'm talking about when I'm saying you can't do anything for its own sake.

Sorry for the form of this reply, I swear I'm not trying to be witty or anything. 

I've been wondering about this, though. Why am I so fired up to find these correspondences? Why are the analogies so fascinating? If I'm using this as a guide for finding likely avenues for meditative exploration, I guess that's a good thing. But if it's leading me to believe that I'm locating some more fundamental truth about the "way things are", that's potentially a problem. I don't think improving communication is the main driver here.

Yes, and this applies directly to how natural selection has gotten us to this juncture. And you make a point that Donald Hoffman regularly makes - we do not perceive the truth, whatever that may be. We perceive that which allows us to survive.

So when Ni Nurta talks about varying his learning rate (let's keep putting words in his mouth until he shows up!), I think what he's trying to say is that he is consciously controlling how quickly his mind adapts to new input. Sometimes going very "fluid", and changing quickly, other times firming up and only changing if new input remains consistent over a longer period of time.

But who knows!

I really like the idea of letting Ni Nurta explain his comments. I don't want to assume we can know what he meant. It's a great mystery - why spoil it?

I'm all ears as well! 

Agreed!

I am not sure what Chris expects but it feels like I was to explain why he feels that me equating meditation to ML feels controversial to him

Anyhoo, Chris, do you know how higher learning rate feels versus lower learing rate?
If not then I recommend getting to know it - useful stuff.

Ni Nurta,

My question was pretty simple, really: Please describe machine learning (how does it work?) and then describe how human beings learn, then compare and contrast the two. 

This seems to be a way to help folks here understand your earlier comment, which I'm certain makes sense to you but isn't clear to others here. I'm trying to have a sincere conversation with you and learn something thereby. I'm truly interested and I hope you can reciprocate and accommodate.

The right way to hear someone talking dharma is experiencing it immediately.
I can do it because I increase learning rate of parts of my neuron network to let them "take it all in". Reducing loss function which is distance between experience they would have while they are not arisen and what it must be to have experience described making actual experiential sense. Then lock it in and can use in the role of "THIS", reference to train rest of the mind and do while it hasn't yet arose so when minds from them arise they are pre-trained. For practical purposes increase of learning rate arises at the same time as noise which is used to reduce loss function. Noise can be very specific to speed things up but it has to always be possible and there should be readiness to use pure random noise.

In general form of it I where it is conscious I call it "mind state visualization". People who know nothing about anything call it meditation and spend considerably longer time to do anything than it would otherwise be needed. Otherwise it is biological machine learning, here optimized for structure of human mind and its strengths and limitations. Main limitation being we are made from beings which can get tired and can be moody. At least we have like ~86 billion of them

Chris, for now I will leave it at that and will maybe pick this topic when I feel your learning rates are not on the level of the Arhat (read: clinging to red hot coal) but more like they are supposed to be at the Stream Entry (read: actually flowing).

In any way "attainment" of Stream Entry means person has always the possibility to drop whatever BS they cling to. What about you Chris, can you drop THIS?

Ni Nurta, I'm truly sorry we can't communicate better about what you post here on DhO. It would be nice to understand you. Alas, it's not to be. I'm dropping this because you won't engage in a real conversation. You want to avoid answering a simple question, make it about me, and respond with ever more obfuscating personal jargon.

I give up. You win. May your neurons rejoice!

​​​​​​​

Not really controversial. Lots of machine learning people do meditation and compare various meditation phenomena to machine learning

I myself mapped out fire kasina phenomena with a model based on convolutional neural networks.

Objectless meditation can be seen as a variant of sampling from a variational auto encoder.

Meditation on an object can be seen as a conditional generative model or classifier.

​​​​​​​People just use whatever metaphors they are familiar with to project their experience onto. It can be quantum woo or machine learning woo it's just language trying to describe the ineffable.