Supercomputer takes 40 minutes to simulate 1 second of a human brain | Hacker News Reader

Supercomputer takes 40 minutes to simulate 1 second of a human brain

tweaktown.com

45 pointskirtijthorat44 comments

Open articleView on HN

Better writeup (cites sources!) and discussion from 5 months ago.

https://news.ycombinator.com/item?id=6157157

And the Riken Lab press release

http://www.riken.jp/en/pr/press/2013/20130802_1/

This is a link to the poster presentation. There does not seem to be a full paper associated with this research yet.

http://www.biomedcentral.com/1471-2202/14/S1/P163

The NEST simulator (The researchers Morrison and Diesman are integral people on this project):

http://www.nest-initiative.org/index.php/Software:About_NEST

I looked up NEST, written in C++ with Python. What, no LISP?

When I hear AI being discussed LISP comes up fairly often. In the book "Meta Maths: The Quest for Omega" by Greg Chaitin, he mentions that Kurt Godel's work had a notation that was uncannily similar to LISP code. On the other paw he compares Turing's idea of code as something more akin to machine code.

Historically speaking, Lisp was the de-facto language of AI for quite a bit of time. That's only been the case for classical AI research, however. People working on AI and machine learning these days aren't necessarily all working in Lisp -- they're usually working in high-performance languages, or in interpreted languages with high-performance libraries (e.g. Python plus numpy or scipy or what have you).

Greg says he wrote his first LISP interpreter in FORTRAN, around the early 1970's. I was thinking that NEST might have code that emulates certain properties of LISP, I've heard of many LISP interpreters being written in other languages over the ages.

> The experiment on simulated human brain activity involved 1.73 billion virtual nerve cells that were connected to 10.4 trillion virtual synapses

So... 1-2 orders of magnitude smaller than a human brain. From Wikipedia:

> One estimate puts the human brain at about 100 billion (10^11) neurons and 100 trillion (10^14) synapses

http://en.wikipedia.org/wiki/Neuron#Neurons_in_the_brain

Assuming linear scaling, that would put an actual simulated second of human brain neural activity somewhere between 6 hours and 2 days.

If it's 2 days that puts us about 26 years (2040) away, compared to the original assumption, which would put us 16 years away (2030), assuming "Moore's Law" continues to happen every 18 months (regardless of whether we'll be able to double the number of transistors per same amount of space anymore or not).

Except Moore's law most likely isn't continuing for much longer, if it hasn't stopped already.

Even if this was actually simulating a proper human brain it's still a silly comparison. Even now a modern computer struggles to simulate an old SNES perfectly at full speed. A SNES is vastly slower than a modern computer but the additional cost of emulating something can be very high indeed depending on how accurate you want the simulating to be. A computer is also very general purpose. I'm sure that some custom built chips and electronics would be much better at simulating these kinds of networks.

I've been running SNES emulators for a decade at full performance. What are you referring to?

He was referring to accurate emulation, rather than the approximate emulator you get with a normal emulator. Here's a related story: http://arstechnica.com/gaming/2011/08/accuracy-takes-power-o...

Probably this: "It doesn't take much raw power to play Nintendo or SNES games on a modern PC; emulators could do it in the 1990s with a mere 25MHz of processing power. But emulating those old consoles accurately—well, that's another challenge entirely; accurate emulators may need up to 3GHz of power to faithfully recreate aging tech." - http://arstechnica.com/gaming/2011/08/accuracy-takes-power-o...

I don't think we would need to simulate a human brain 100% to have strong artificial intelligence. Similarly how we don't need to simulate the SNES 100% to be able to enjoy the games.

Furthermore I'm pretty sure a human brain also maintains various bodily functions, which use some percentage of the brain's computational power.

I also suspect that strong AI is probably not as complicated as we think, it's just that no one has thought of the correct set of ideas required for strong AI to emerge.

Great, we're only 11 doublings away. By Moore's Law, that means we should have AI ready just in time for it to fix the Year 2038 problem.

it is still hard pressed to compete with the brain in your head reading this article.

It took K around 40 minutes to simulate just 1 single second of human brain activity, even with all of its performance prowess. The experiment on simulated human brain activity involved 1.73 billion virtual nerve cells that were connected to 10.4 trillion virtual synapses, with every virtual synapse containing 24 bytes of memory.

There's no way the brain in my head could simulate 1.73bn nerve cells in 40 minutes.

We still can't properly simulate the puny caenorhabditis elegans brain with some accuracy. Don't hold your breath on this one.

This. And we've had the complete wiring diagram for its brain for over 20 years.

There's a wide-spread assumption that knowing the brain's connectivity will be sufficient to emulate its function, but there's so much about the underlying molecular and electrical properties of neurons that we know nothing about.

It's somewhat analogous to having a circuit diagram where the components are missing. Dropping in random components but preserving the wiring structure won't result in the same functionality.

An ACTUAL simulation of rat brain would be 1000 000x more useful.

And would carry a lot fewer disturbing implications about mistreatment of simulated former neurological patients.

If they really were simulating a human brain, that would have some pretty serious moral implications.

I think in the short term, companies that develop these things will behave ethically without any oversight (by making machines "enjoy" what they are doing) because doing something else would be inefficient or counterproductive. Why would you make an expensive thinking machine miserable? Humans that are happy are way more productive - and machines that are based off humans will be as well.

In the long term, if and when these things become mass-produced and cheap, people may want to do terrible things to them, in the same vein as animal torture. That may be when laws get put in place to protect them.

Ethically? I bet they will kill them thousands of times during development.

Suppose you, at one stage, have a simulation of a brain that isn't quite there; it talks and sees, but it's audio system doesn't work right. What do you do?

Even live debugging to repair it can be controversial (http://en.wikipedia.org/wiki/Cochlear_implant#Criticism_and_...)

I don't see anything unethical about shutting it off. If nobody is emotionally attached to it, and it doesn't suffer when it is shut off, who is harmed by the shutdown?

Consider a person without relatives or friends. Would you consider it ethical to shut him/her off, as long as that person didn't suffer from it?

As long as they do not care that they could be "shut off", I see nothing wrong with it. If they dislike that notion (like real humans do), then the possibility of shutdown would cause suffering and would be immoral/unethical to allow.

You're assuming that the machines will care about being shut off - we would probably design them so that they don't care about this, because this makes them easier to work with. And then it's no longer unethical.

I don't know. If you can simulate a brain, you can alter it. And if you can alter it, you can make it artificially happy, or simply remove areas hosting willfulness, sleep, sexuality or independent thought. Won't make them suitable for all tasks, but for some it would be more efficient.

Ethical? That's the question.

Suppose the simulated mind goes insane within a few seconds. Just long enough to solve a captcha, before being restarted...

Not at all. Simulating your brain != simulating your mind. Don't become a victim of 'neurobollocks'. http://www.newstatesman.com/culture/books/2012/09/your-brain...

Your New Statesman link only talks about pseudoscience and doesn't support your assertion. The mind is an emergent property of the brain. If you can't simulate a mind by simulating a brain, then what in your opinion would it take? Or is it impossible because we all have ineffable souls that the laws of physics mysteriously cannot access?

> The mind is an emergent property of the brain.

This is the kind of pseudo-science that gets so much attention nowadays. Incredible.

This is exactly the sort of problem that convinced me that consciousness/mind must be a fundamental part of the (multi/uni)verse, and not something that "emerges" from matter.

I find it terribly unconvincing that a specific arrangement of matter, or even an algorithm in software, just simply "spawns" a discrete consciousness from nothing. It might make sense if there was an underlying "consciousness field" or some such concept that the matter-arrangement tapped into in some manner.

If we assume that, could a machine then tap this field?

> I find it terribly unconvincing that a specific arrangement of matter, or even an algorithm in software, just simply "spawns" a discrete consciousness from nothing.

It's not only unconvincing but unscientific, a pseudo explanation.

One thing we have to understand here is that 1 second of human brain activity is quite a lot of computing! It's not 'just one second'. This incredible experiment shows how the human curiosity has gone so far to build a artificial thinking machine. It would be great to see the compute result of if we could add 10 Supercomputer.

If we simulate a human brain as a way to make computers solve new types of problems, then we will have bored computers who procrastinate solving problems in favor of playing WoW.

What does simulating the human brain mean? Would a simulation lead to artificial intelligence or just quite a large neural network?

Seems like just a large neural network (which could be of coursed used for AI purposes).

More detail is here in the original press release: http://www.riken.jp/en/pr/press/2013/20130802_1/

Hmmm... I wonder If we could record the state of an active brain and "run" it in the computer creating sentience. Without getting theological, what would the difference between the brain and the computer be if they were exact copies and functioned in the same way?

That's one of the many subjects of Ghost in the Shell. From [0], let your web-fu walk the path.

[0]: http://en.wikipedia.org/wiki/Philosophy_of_Ghost_in_the_Shel...

The problem currently is that we actually don't know how the brain works. I know a lot of people claiming that they've found out - but if you dig deeper you'll find most experiments only work in certain cultures and are far from being THE theory which explains everything. I mean nobody today can actually be sure that the power at which a synapse is firing has a meaning. We simply have no idea - it may have significant influence but we're far from proving it (can't find the paper describing this, as I'm not able to connect to the university network currently).

This is bad news - especially for the AI-robotics guys, as they need this knowledge to implement the next generation of smart robots. They hope to get some "self reprogrammable robots" as this is what your brain seems to do all the time.

So what do we technicians do? We're trying to build a machine to simulate a brain (and hope we're right in our assumption how the brain works). There is a huge project like this going on in the EU too [1]. This is the bottom-up approach and it's far from all the press releases as there are too many assumptions in it - even if those guys hate to hear it. There is a nice documentary film with Jospeh Weizenbaum (former professor at MIT and close friend of Chromsky) about this very issue and its ethical aspects [2].

The up-down approach is researched by system biologists (and other related sciences). They aim at the bio-chemical and physical layers to figure out how a brain works and it seems like this is complicated as hell. We're some kind of programmable - even if nobody can tell how far this actually goes. Just being raised in different cultures can have significant influence in how a brain reacts in situations. Even siblings with the same DNA and are being raised within the very same family you can find differences in how their brain reacts...

So don't get too excited about all this - we're far away from being "downloadable". Which is maybe not the worst thing if you're thinking about it for a while. What would life mean if it's endless? And by the way: even if you were downloadable, what tells you that you're still alive if a copy of your brain is stored within an robot? There are hard philosophical questions behind such issues...

[1] https://www.humanbrainproject.eu/ [2] https://en.wikipedia.org/wiki/Plug_%26_Pray

Well isn't the real hope that we'll get the hardware close enough in capability that one day a learning general AI just kind of pops out of it?

Does this mean we are ~216 months away from having 1:1 performance?

Does this mean we are ~100 months away from 1:1 performance?

Actually a shorter amount of time than I was expecting.

This article from last summer speaks of 1% of human brain activity:

http://news.cnet.com/8301-17938_105-57597049-1/fujitsu-super...

It also says "The synapses were randomly connected" among other more even handed discussion of what happened.

Reply on news.ycombinator.com