I'm a neuroscience major, and I'm curious: should we have a greater focus on emulating the design of the brain in our design of computer systems?
One obvious advantage of the brain is power consumption (not relative to the body but comparatively to computer systems).
We also obviously parallel some aspects of the brain in terms of neural network design (I see a lot of parallels of convolutional neural networks and the idea of "grandmother cells" (more accurately the work on Halle Berry/Jennifer Aniston cells)
But i think it goes further than just that. What about microkernel design instead of monolithic kernels? It seems to have parity with the concept of conscious and unconscious control of body functions in humans.
There's also the idea of concurrency, as we're getting close to reaching the end of moore's law, so I feel we're going to have to move to programming and architectural environments where concurrency is favored rather than relying on raw processing speed (eg maybe we'll have 128 core computers in 5-10 years), similar to how the brain has "specialized sections" that work in parallel and share information with each other.
It just seems like a very interesting idea to me that could be explored further than just the neural network paradigm all the way down to basic architecture. What do y'all think? ( I would post on /g/ but they're less inclined towards interdisciplinary topics)
in a way Siri already works this way. You have a device which can capture sound (your phone), that sound gets sent to a server where it's analyzed and converted to text, the text is then sent to another server where it's analyzed for instructions or what Siri is suppose to be doing, then those instructions are formatted in a way your phone can understand, then all that info is sent back to your phone again where it can execute the instructions.
Jacob Powell
First impression is no just because we can obviously see brains are designed to do something come tell different compared to your average computer. And then I guess we don't really know enough.
Mason Edwards
Society has only a tiny idea of how the brain works: It's hard to mimic something when one does not understand that something. Are better computers going to do "anything" to solve societal problems? Unlikely. It's more likely computers are now the problem: people don't "think," and this has always been a problem.
Chase Clark
Ok but think of this:
It takes a tiny interval of time to analyze a face and say "this is Halle Berry". What if we used that same structure but with different algorithms to solve, say, complex math equations (specifically ones that could benefit from concurrency)? This kinda is just aping the idea of neural networks, but it seems like anything that can be done concurrently could be modified into a brain like system. Beyond that, we know drugs can temporarily alter the connections we make, so it seems like brain-like systems are very open to metaprogramming, allowing us to alter our systems on the fly for a variety of different problems.
Jose Edwards
Yeah but I don't care about society's problems. I care about making efficient systems, which is a prerequisite for doing a lot of advanced computation that we can't do right now, eg creating an artificial intelligence that is more than just statistics based. Also we know a shitton more about the brain than you would think. We can pinpoint a cluster of neurons linked to your perception of a specific actress, for example. That's crazy.
Bentley Phillips
Grandmother cells are a bit of a meme though, at least I was left with the impression that decentralized thinking and memory is the accepted model.
>creating an artificial intelligence that is more than just statistics based aren't we basically statistics based too?
Blake Lewis
They're a "bit" of a meme, yes, in that it's not literally one cell, but it's small clusters, so not really true decentralized thinking. True grandmother cells are just an extreme form of reality.
And yes we are kind of just statistics, but I mean abstracting past that level to a point where they're able to function on a practical level like humans.
Benjamin Campbell
Brains are kinda shit at a lot of things computers are good at though. I wonder how much it would help. Depends what you want the brain to do I guess. Computers are designed for like calculations and shit aren't they. I've seen brains described as like generative models in statistics and the inversion of those models in producing perception and action as predictions of the internal, external and motor (muscle,viscera,autonomic) environments and that's essentially what they do. We also forget that it takes a relatively long time for the brain to learn stuff even so, they aren't preprogrammed.
Maybe will be useful. We don't know enough to know if it would be interesting think. I don't think it would make computers necessarily more efficient. Maybe machine learning.
To be fair we don't know too much about grandmother neurons. I suspect if they do represent e.g. Jennifer anniston, they only represent it in the direct context of the task and still rely on distributed inputs from elsewhere so still quite complicated but would be interesting to read a paper with that.
Isaac Cooper
Machine learning I guess that kinda stuff maybe.
Cameron Turner
>they aren't preprogrammed This is false. A significant amount of information is "programmed" into the brain from birth. For example all the automatic body processes handled by the brainstem like breathing or heartbeat, and if you throw a young baby into water it can swim even without ever seeing somebody else do it. Basic motor functions are also preprogrammed, and some more abstract things like being scared of certain shapes and sounds but comforted by other shapes and sounds
Luis Scott
Yeah of course there's much innate but i just mean pretty much anything complicated or useful that you might want a computer to do in general takes a long time. Even normal vision for instance requires experience. I feel like the majority of innate preprogrammed ability will be mediated subcortical as opposed to say the prefrontal areas where the most complicated stuff happens that takes long time.
Cameron Collins
Once we understand how the system of memory storage works we could presumably come up with a way of converting conventional information into the "brain" style information quickly and directly
Also brains are definitely programmable, on many levels. LSD temporarily reprograms the brain, nootropics do too, brainwashing is also real and etc.
And think about how we combine various traits to say "that's a face". Maybe we could use that same method but with a different algorithm to quickly factor numbers, for example.
Jordan Price
We don't know how its preprogrammed though.
Robert Wood
I feel like maybe the issue is computational power rather than specific things about the brain. I think we do know vaguely how it brain works just very very complicated.
Brody Bennett
Well there's one easy reason : it's difficult to build stuff in 3d. A computer chip is mostly a 2d layer.
Some people are trying to build multilayer chips, but that's difficult, because reasons.
Now the brain is just way more complicated than that, because the links are not even layered, they can go every fucking direction and distance. So reproducting the brain basic architecture is not something that will happen soon. Now every one knows that distributed systems are the future of computing, except many problems arise there as well, see en.wikipedia.org/wiki/Category:Distributed_computing_problems
The basic idea, is that since things are distributed, well, they are not united. If one component tells you A and another component tells you B, what do you do ?
Christian Williams
Is neuroscience a good undergrad major if I want to do AI in grad school?
Justin Moore
No, study math or CS if you can't handle math. They focus on similar problems, but neuro and AI are substantially different in process and what you learn to apply.
Caleb Martinez
...
Luke Phillips
How do nootropics "reprogram" the brain? Are they helpful for conditioning? Any examples or papers on this?
