Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Neurons in a dish learn to play Pong

nature.com

Post reply on HN
Print view View original

81–90 of 144 posts

Re: Neurons in a dish learn to play Pong

#81

Post by rryan »

lern_too_spel wrote: post #71
d--b wrote: post #64

> he work is a proof of principle that neurons in a dish can learn and exhibit basic signs of intelligence Absolutely not. This shows that you can teach neurons to exhibit a reflex behavior adapted to the given stimuli - which shouldn’t come as a surprise. At best this is jellyfish-like level of intelligence.

Transformers have shown that language development can also be explained by what you derisively call reflex behavior, with no need for specialized Chomskyan recursion circuitry, as the same mechanism with slight modification (decision transformers) also plays Pong.

There's a huge difference between fitting a probabilistic model to a data distribution then sampling from it (what GPT-3 is) and agents that invent language and use it to communicate.

Re: Neurons in a dish learn to play Pong

#83

Post by leonmoonen »

sillysaurusx wrote: post #79
thrown_22 wrote: post #78

Earlier quoted context omitted.

Do you know the actual study which did this? Because the Nature citation on where the work was done is literally "The Discovery Channel".

It was big news at the time. Headlines everywhere. Rat brain flies plane. You’d think that within a few years, maybe rat brains would be the primary way we’d control planes. I did manage to track down the original study, which at the time wasn’t too easy; this was before scihub. I wish I’d saved it for posterity. My motivation went from “this is exactly what I want to do with my life” to laying on the couch wondering…

Thomas DeMarse is mentioned in some of the 2004 news articles. His Google Scholar profile has the following short paper published at a conference in 2005: https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=Thom...

Re: Neurons in a dish learn to play Pong

#84

Post by justin_ »

sillysaurusx wrote: post #79
thrown_22 wrote: post #78

Earlier quoted context omitted.

Do you know the actual study which did this? Because the Nature citation on where the work was done is literally "The Discovery Channel".

It was big news at the time. Headlines everywhere. Rat brain flies plane. You’d think that within a few years, maybe rat brains would be the primary way we’d control planes. I did manage to track down the original study, which at the time wasn’t too easy; this was before scihub. I wish I’d saved it for posterity. My motivation went from “this is exactly what I want to do with my life” to laying on the couch wondering…

I think it might be this one:

Adaptive flight control with living neuronal networks on microelectrode arrays

The brain is perhaps one of the most robust and fault tolerant computational devices in existence and yet little is known about its mechanisms. Microelectrode arrays have recently been developed in which the computational properties of networks of living neurons can be studied in detail. In this paper we report work investigating the ability of living neurons to act as a set of neuronal weights which were used to control the flight of a simulated aircraft. These weights were manipulated via high frequency stimulation inputs to produce a system in which a living neuronal network would "learn" to control an aircraft for straight and level flight.

https://ieeexplore.ieee.org/document/1556108

https://sci-hub.ru/https://doi.org/10.1109/IJCNN.2005.155610...

Re: Neurons in a dish learn to play Pong

#85

Post by nier »

seydor wrote: post #4

i would urge people to read the paper instead. The 'learning' is a bit iffy , and this was meant to test the brain theory of Friston rather than plug neurons into pong. Still, great job on the neurotechnology involved and a step in the direction where we should be going, controlling large numbers of neurons Ars Technica has a better article, although they dont describe the dense electode array correctly: https://arst…

For those who like me are wondering where those neurons come from. An answer can be found in the Ars Technica article:

“[...] the researchers tested two types of neurons: some dissected from mouse embryos, and others produced by inducing human stem cells to form neurons. ”

Re: Neurons in a dish learn to play Pong

#86

Post by bowsamic »

nazgul17 wrote: post #75
d--b wrote: post #64

> he work is a proof of principle that neurons in a dish can learn and exhibit basic signs of intelligence Absolutely not. This shows that you can teach neurons to exhibit a reflex behavior adapted to the given stimuli - which shouldn’t come as a surprise. At best this is jellyfish-like level of intelligence.

Every new AI gets demoted pretty quickly. Playing pong would have been once called AI.

That’s because once we understand it and achieve it it’s no longer considered intelligence. The same will happen when we have actual human robots that are indistinguishable from real humans. People will say “obviously not intelligence or anything related to it, just doing x, y, and z”

Re: Neurons in a dish learn to play Pong

#87

Post by jamiek88 »

lvass wrote: post #66
sidewndr46 wrote: post #5

A question for someone who understands the neurology and biology of this much more than I do: Once I have a group of neurons like this trained to do something, can I actually count on them to continue performing that task until they die? Or is it possible they spontaneously reorganize or "learn" a previously unseen behavioral pattern?

A related question: can I indefinitely test they're performing their task correctly? When operating these neurons for actual tasks, real time challenge-response-like testing is probably a good idea. But if they can learn, eventually they may be able to trick the tests. How would one design automated tests to evaluate the state of learning-capable systems?

>How would one design automated tests to evaluate the state of learning-capable systems?

A question many call center managers ask themselves.

Re: Neurons in a dish learn to play Pong

#88

Post by sillysaurusx »

justin_ wrote: post #84
sillysaurusx wrote: post #79

Earlier quoted context omitted.

It was big news at the time. Headlines everywhere. Rat brain flies plane. You’d think that within a few years, maybe rat brains would be the primary way we’d control planes. I did manage to track down the original study, which at the time wasn’t too easy; this was before scihub. I wish I’d saved it for posterity. My motivation went from “this is exactly what I want to do with my life” to laying on the couch wondering…

I think it might be this one: Adaptive flight control with living neuronal networks on microelectrode arrays The brain is perhaps one of the most robust and fault tolerant computational devices in existence and yet little is known about its mechanisms. Microelectrode arrays have recently been developed in which the computational properties of networks of living neurons can be studied in detail. In this paper we repor…

Thank you so much! That’s the exact one. I’ll upload it to imgur, since scihub urls tend to die over time: https://imgur.com/a/vdQP17I

The details are presented at the end of page 2. You can see on page 3 that the whole thing is a glorified “I can make neurons go to zero!” machine.

It kills me because there is real value in this work — the ability to modify neurons is a useful thing to be able to do. Showing that you can set them to specific values is worthwhile.

But that wasn’t what it was presented as. The focus was on the idea that the neurons somehow learned something. Maybe they did, and maybe time will prove me a fool by showing that this is how neurons learn in general. But it was so disheartening to spend hours carefully going over every detail, full of excitement at the possibility of machines that can learn… only to realize I could do the same thing by setting a value to zero in python, and that the complicated language seemed designed to conceal this.

I remember seeing a video on the discovery channel where they were interviewing him. He gave a dazzling account of the implications of this work.

Maybe he was predicting the upcoming ML boom. But now that I’ve worked in ML for a few years, I can safely say that this work didn’t help us get here. Maybe it helped biologists figure out how to control the signals that neurons deliver.

(I still think that it’s pretty darn cool that you can manipulate neurons at all, so hats off to whoever figured out how to do that. I just wish it was presented as what it was: the ability to set a value to zero over time via neurons.)

I really want to make a joke like “You know what else would set neurons to zero? Set them on fire and wait,” but the paper showed they could be set to arbitrary values. Which of course was the interesting and valuable part in the first place, not the plane stuff. But planes make good headlines.

Re: Neurons in a dish learn to play Pong

#89

Post by mattlondon »

d--b wrote: post #64

> he work is a proof of principle that neurons in a dish can learn and exhibit basic signs of intelligence Absolutely not. This shows that you can teach neurons to exhibit a reflex behavior adapted to the given stimuli - which shouldn’t come as a surprise. At best this is jellyfish-like level of intelligence.

Agreed....but! This opens some interesting thoughts about various things.

Sure this Petri dish is not intelligent and conscious. But it's just a load of neurons right, just like the Petri dish in our head. Are we intelligent and conscious in that case? Or are we just a scaled up Petri dish that reacts to more varied stimuli?

Re: Neurons in a dish learn to play Pong

#90

Post by seydor »

ummonk wrote: post #43

It seems needlessly provocative to use human neurons to do this. Presumably they could have had the same effects with neurons from another species. It’s like they want to provoke a backlash from religious people against scientific experimentation.

technically they used human skin cells, which they induced to be pluripotent, which then became neurons
Post reply on HN