How deep is the brain? The shallow brain hypothesis

> we're in the same situation with animals and people think that, what, understanding gradient descent or backprop is helpful? this is just some csci bs

Assuming I've actually got your point for this (and I'm not sure I have):

The backpropagation algorithm itself might be "just some csci bs" (it sure has vibes of "let us shortcut the maths rather than find out how our brains did it"), but gradient descent is nice and general-purpose — much like how evolution is both good for biology and in simulation for everything else.

Now, try to take that seriously and think about the laptop as an actual object of experimental curiosity -- what exactly does science need to invent, discover, describe etc. to understand the operation of that laptop?

99.999% of that new knowledge has to be in physics and chemistry, before the tiny 0.0001% of theoretical csci knowlegde is brought to bare.

Consider how impossible it would be to apply any csci knowledge first: we do not even have the ability to measure the cpu state! So we could not even identify any part of the system with 0s, 1s, etc.

Now: that's a laptop!

Imagine now you're dealing with an animal.

Hopefully its now clear how ridiculous it is to describe basically any aspect of our mode of operation by starting with trivial little csci algorithms. It would be insane even with an actual electronic computer, let alone an organic system.

A system whereby clearly our organic properties are radically fundamental to our mode of operation

On the contrary. That's an implementation detail. You can "power devices, and modulate devices" by having a clockwork computer with transducers at the I/O boundary, converting between electricity and mechanical energy at the edge. It would work exactly like a fully electronic computer, if built to implement the same abstract computations - and as long as you use it within its operational envelope[0], you wouldn't be able to tell the difference (except for the ticking noise).

> The causal properties of matter are essential to any really-existing system. Non-causal, purely formal properties of systems which can be modelled as functions from the naturals to the naturals (ie., those which are computable) are useless.

Yes and no. Of course the causal properties of matter... matter. But the breakthrough in understanding, that came with development of computer science and information theory, is that you can take the "non-casual, purely formal" mathematical models of computation, and define some bounds on them (no infinite tapes), you can then use the real-world matter to construct a physical system following that mathematical model within the bounds, and any such system is equivalent to any other one, within those bounds. The choice of what to use for actual implementation is done on practical grounds - i.e. engineering constraints and economics.

It's how my comment reached your screen, despite being sent through some combination of electrons in wires, photons down a glass fibre, radio signals at various frequencies - hell, maybe even audio signals through the air, or printouts carried by pidgeons[1]. Computer networks are a living proof that substrate doesn't matter - as long as you stick to the abstract models and bounds described in the specs for the first three layers of ISO/OSI model, you can hook up absolutely anything whatsoever to the Internet and run TCP/IP over it, and it will work.

I bet there's at least one node on the Internet somewhere whose substantial compute is done in a purely mechanical fashion. And even if not, it could be done if someone wanted - figuring out how to implement a minimal TCP/IP stack using gears and switches is something a computer can do for you, because it's literally just a case of cross-compilation.

--

[0] - As opposed to e.g. plugging 230V AC to its GPIO port; the failure modes will be different, but that has no bearing on either machine being equivalent within the operational bounds they were designed for.

[1] - https://datatracker.ietf.org/doc/html/rfc1149

Consider two hypothetical versions of this. One, the exact scenario as you described - history unfolded like it did, until the 1900 alien incident. CS and information theory is in its infancy. You're correct that most of the necessary work would first go to physics and chemistry and their various spin-off fields, because that's what's needed to build tools necessary to inspect the machine in full detail. The math would develop along the way, and eventually enough CS to make sense of the observations made before.

Now for an alternate scenario: it's the 1900 again, with the twist that CS is already well-developed theoretical field of mathematics (IDK, perhaps the same aliens dropped us a mechanical computer in year 1800). We'd still need to push physics and chemistry (and spin-offs) forward, but this time, we would know what we're looking for. We'd know the thing does computation, we'd be able to model what kind of computation it does. The question would change from "what does this thing do" to "how exactly does it compute the specific things we know it does". I imagine this would speed up the process of getting a complete picture, because it's easier to understand a specific solution to a problem once you know the answer, than it is to figure out the answer along with the solution.

In terms of understanding the brain, we are in the second situation. We may still know little about how the gooey thing ticks, but we have a growing understanding of what comes out of all that ticking, and a very good understanding of the fundamental rules of ticking.

Neurons are not connected by a simple graph, there are plenty of neurons which affect all the neurons physically close to them. There are also many components in the body which demonstrably affect brain activity but are not neurons (hormone glands being among the most obvious).

> with weighted connections

Probably, though we don't fully understand how synapses work

> performing distributed computation (mainly hierarchical pattern matching)

This is a description of purpose, not form, so it's irrelevant.

> learning from data by gradually updating weights

We have exactly 0 idea how biological neural nets learn at the moment. What we do know for sure is that a single neuron when alone can adjust its behavior based on previous inputs, so the only thing that is really clear is that individual neurons learn as well, it's not just the synapses with their weights which modifies behavior. Even more, non-neuron cells also learn, as is obvious from the complex behaviors of many single-cell organisms, but also some non-neuron cells in multicellular organisms. So potentially, learning in a human is not completely limited to the brain's neural net, but it could include certain other parts of the body (again, glands come to mind).

> using selective attention (and/or recurrence, and/or convolutional filters).

This is completely unknown.

So no, overall, there is almost no similarity between (artificial) neural nets and brains, at least none profound enough that they wouldn't share with a GPU.