[mjburgess]

There is no biological analogue to backprop, any supervised learning step, etc. Nor, conversely, a mathematical analogue of neuroplasticity, biochemical signalling, etc.

I haven't gone far on looking at CNN vs. visual context, but I could only imagine the analogue exists at the hierachical-geometric level; ie., it isnt a model of function, but simply a model of how one can in principle process visual information.

(Which any system parsing information would follow, AST parsers, etc. are likewise hierarchical.).

[zcw100]

I find this article that I recently read very relevant to the discussion "On the cruelty of really teaching computing science" for anyone not familiar with the article I'll leave it to anyone reading to wait until they get to the end.

https://www.cs.utexas.edu/users/EWD/transcriptions/EWD10xx/E...

[Der_Einzige]

Forget the biological connection, I'm sitting here waiting for someone to write a paper about how gradient descent and backprop are basically equivalent to the platonic notion of the dialectic within epistemology.

I wouldn't throw out biological plausibility, especially with the spiking neural network variant of them...

[datastoat]

While we're waiting for the paper on the Platonic link ... here's a book about how machine learning is all just a rerun of Oliver Wendell Holmes's theory of epistemology in the law, from the 1890s.

https://link.springer.com/book/10.1007/978-3-030-43582-0

[monocasa]

For some biological neural nets, one would imagine the loss function just being nearly completely genetic. That is, both the weights are ultimately stored in the genetic code the same way any other cell differentiation information is stored, and were converged on in the first place by random mutations being more or less fit for their ecological niche. Particularly this could explain the relatively fixed function components like our V1 and V2 regions, our motor region, and probably the entire nervous system of simpler animals.

Basically in a lot of cases evolution takes the place of back proposition, and large structures are what we'd call inference only without learning. Doesn't make them any less of neural nets.

[amelius]

There's this theory, often cited in this context:

https://en.wikipedia.org/wiki/Hebbian_theory

[mjburgess]

It's not clear how this sort of ideas stack up against modern critiques of computational approaches.

Eg., this sort of analysis has been applied to CPUs where it renders incoherent/obviously false results.

[salawat]

>There is no biological analogue to backprop, any supervised learning step, etc.

Excuse me? How do you figure? Have you tried to do any physical task before, and not gotten it right the first time, tried again, and got closer? Hello, backprop.

Has someone trained you, telling you you're doing it wrong? Hello, supervised learning. There is an entire portion of your brain tasked with measuring the difference between what you intended and what you did. If you damage it, you actually end up in a state where other people's feedback is what you have to rely on.

The equations behind NN's are "non-biological" only in the sense that you're taking a general math function (the sigmoid) and burning it into silicon gates that are not part of a living organism that also has to deal with the excess baggage of remaining alive, or operating as a chemically based computer.

However, the dynamics and fundamental capabilities of a biological neural network and a silicon based one are fundamentally the same. You can just scale input domains and speed of the Silicon based one a lot easier than you can the bio-based one due to the difference in computing media. You can also wipe out and retrain a silicon net without being considered to have "killed" anything. The Silicon nets, interestingly, have no analog of running and training simultaneously, or are just starting to get it from the lit I've kept up with, nor an analog for forgetting typically applied to them, making the biological net the far more interesting information processing construct. To be frank, NN's not resembling the biological models says more about the inefficiencies of our media of computing, our woeful lack of understanding around the nature of human/biological perception, and the arrogance of human beings who lack the capacity to recognize in the math something fundamental to their own existence.

This post written by a multiple decade uptime GA/C/R/D pick your letter, constantly training internetworked lattice of sigmoid encoded control networks whose morning your lack of perspective just tainted, and whose training is thusfar unsuccessful in attempts to extinguish the need to point out mis- or incomplete understandings of those sharing the same messaging medium as it generally does. Your overfitting to textual and numerical symbolic representation recognition will not serve you well if you can't extend that capacity to discern the pattern into the overaching context of human existence. That's where you really start gaining an appreciation for the power of NN's. It's the building block of the information processing construct that can eventually recognize and reproduce itself.

Now if you'll excuse me, I need to get back to some obligatory biological maintnance. This bag of organs doesn't maintain itself, you know.

[mjburgess]

Distinguish a rock rolling down a hill from a brain: a rock rolling down a hill finds the best path by having its route supervised by the surface of the hill, etc.

Everything is everything if your level of abstraction is "thing".

The relevant characteristics of "biological intelligence" and of the functioning of the brain do not exist at the "try and try again" level of abstraction. (At this level, as above, we couldn't distinguish and animal from a rock; nor, I imagine, basically any physical process from any other.)

Organic systems grow in response to interaction with their environments, acquiring novel physical structure and causal properties. Neurological intelligence allows for theory-formulation on single-example cases (eg., a child burning their hand once is sufficient to build a theory of their immediate environment).

The list goes on.

The capacities of these systems do not obtain in the machine case, and likewise, the machine cases has no functional analogues at the relevant level of distinction.

This dumb form of statistics called, "approximate associative modelling over 1tn cases", ie., Machine Learning, has nothing new to say about intelligence, biology or neurology.

We have been doing non-linear regression and optimisation since the victorian era.

[stickfigure]

Sure, you can find arbitrary similarities in everything. You can also find arbitrary differences in everything.

Experiments like GPT3 do seem to point in the direction of "scale" as the dominant factor. Until we can reach the same level of scale as a real brain, the question of whether "meat is special" is undecided. Everything that is you may just be a non-linear regression on a chemical computer.

[mjburgess]

GPT3 isn't even in the same domain as intelligence.

Statistical patterns in trillions of examples is a sideshow.

Words in natural language refer to the world; that is their point. Communication is the coordination of that reference between speakers in a shared environment.

You cannot take GPT3 to new york and ask it what it thinks of the city: it cannot be anywhere (it isnt causally connected to an evnironment); and it cannot coordinate with any listener (it has nothing to say).

Text generation is certainly reaching new heights. This isn't a form of communication, however, and isnt even relevant to it.

[stickfigure]

I think you're falling into the Chinese Room fallacy. I agree that GPT3 isn't sophisticated enough to be considered AGI.

On the other hand, based on the progression of GPT -> GPT2 -> GPT3, remarkable things happen when you add orders of magnitude more nodes to the network.

You might try to argue that no matter how convincing GPT50 passes the Turing test, it's still not intelligent. How is that different from saying the Chinese Room doesn't speak Chinese? Why is your meat-based Chinese Room special?

[mjburgess]

It isn't meat-based, it's in the world.

It's a distinction in kind, not degree. You're presuming that we are just bleak repositories of trillions of sentences stiched together: we arent a meat version of any ML program; not GPT or any other.

We do not learn the meaning of "Green", or "Tree" nor any basic concept via examples in language.

An infinite amount of complexity considering an infinite amount of text cannot refer to the world; it has never been in it.

We aren't statistical patterns in trillions of books. You already presume that GPT is something that it isn't when you presume it is even capable of communicating anything.

[dnautics]

> Have you tried to do any physical task before, and not gotten it right the first time, tried again, and got closer? Hello, backprop.

Backprop is a very specific mathematical operation using derivatives, not a general concept. IIRC there is counterevidence against that neurons train by taking stored inbound activations and updating their weights with a backwards propagating set of gradients.