back to article Boffins build billion-synapse, three-watt 'brain'

A top-notch supercomputer can beat humans at many things, but they're also energy hogs – which is one reason so much work goes into neural simulation as the basis for computers. Now, Stanford scientists have demonstrated a neural simulation which they say has a million neurons, multi-billion synapse connections – and runs on …

COMMENTS

This topic is closed for new posts.

A brain is not the answer.

People of yore used to do calculations by hand, taking years to calculate tables etc. Due to their foibles, they frequently made errors and would have to redo months of calculations.

When it comes to doing the stuff super computers do: sheer calculation, then even 8-bit microprocessors running on less than 1mW do a better job than a brain.

2
15
Boffin

Re: A brain is not the answer.

Well, yes. But no-one is intending to shoe-horn this into doing the sorts of tasks that computers today do well. It is aiming for all the things they do badly.

This isn't modern politics - it isn't either-or!

22
0
Silver badge

Re: A brain is not the answer.

Correct, brains are not that reliable for calculations. On the other hand they somehow manage to very quickly find a 'good-enough' solution to vastly complex more problems than a current supercomputer can handle, and using a tiny fraction of the energy.

This isn't a replacement for computers but a complement

14
0
Silver badge

Re: A brain is not the answer.

Active research is also going in alternative direction, by simulation of human neurocortex

1
0

Re: A brain is not the answer.

As previously posted, the point isn't to replace digital computation, but to hook into the aspects of the brain that can beat computers hands down. Good Enough is the key here.

Besides that, the human brain is inefficient until the particular functions make their way into the subconscious (S of the Rassmusen's SRK model). For example, if I ask you what day of the week a particular date was, any date past or future you will have to work it out with R or K (using rules or knowledge of *how* to do it for calculation) and may get it wrong, but there is at least one person alive today who will be able to just tell you, just like that, 100% accurately, with no conscious calculation. He's using the S part. (Point of interest is the person to whom I'm referring has autism, a largely black hole in our knowledge of the brain). If we can combine what Stamford have done with the subconscious abilities of our brain for bigger problems... wow. I'm excited by this as my final year at uni was looking at neural nets for application in disabilities. It was the most engaging EE and CS project I ever worked on.

5
0

Re: A brain is not the answer.

There's a thought. Design to an autistic human model.

2
0
Silver badge

Re: A brain is not the answer.

A brain is a massively parallel pattern matching system. It also has a relatively ingenious lossy compression memory as well that manages to overlay many memories over each other but somehow keeping them intact enough to be separate.

Any individual, small part of the brain can be implemented faster in silicon, however the ability to form and reform a wide mesh of analog connections between many neurons (quite flexible processing "cores") is something that is very costly to implement.

Repetitive, exacting processes are ideal for procedural computers, however other processes such pattern matching, approximation and detail substitution are much more suited to neural networks. While one can substitute for the other in most circumstances, it is far less efficient.

1
0
Anonymous Coward

Re: A brain is not the answer.

A brain is the answer when you want answers that aren't 0 or 1, yes or no.

Some systems need to be trained and taught, but do this once and you can easily replicate the data meaning the "expert" can be cloned.

0
0

“you have to know how the brain works to program one of these”.

So, how many people that use a brain daily know how it works?

7
0
Meh

Re: “you have to know how the brain works to program one of these”.

How many people use a brain? At all?

12
0
Gold badge
Facepalm

Re: “you have to know how the brain works to program one of these”.

How many people program brains daily? Except for a few films where the villian may be stroking a white cat, living in a volcano, I've not seen any evidence of people programming brains.

1
1

Re: “you have to know how the brain works to program one of these”.

theres always the yanks with their MK-Ultra "experiments".

though the argument could be made that they got it badly wrong...

2
0
Silver badge
Thumb Up

Re: “you have to know how the brain works to program one of these”.

People are continuously programming their own brains and others'. It's called learning and teaching.

12
0
Silver badge

Re: So, how many people that use a brain daily know how it works?

It works on a delicate balance of caffeine and alcohol.

6
0

Re: “you have to know how the brain works to program one of these”.

@Tromos

>So, how many people that use a brain daily know how it works?

How many people that use a toaster daily know how it works?

0
0
Coat

Re: “you have to know how the brain works to program one of these”.

I've seen plenty of instances of cats programming human brains ....

3
0
Silver badge
Pint

Re: “you have to know how the brain works to program one of these”.

"How many people program brains daily?"

Plenty: Teachers, parents. Its their job. It's called "teaching".

2
0

Re: “you have to know how the brain works to program one of these”.

Not really a valid point: you don't have to know how a toaster works to use it.

I think what Tromos is getting at is that the creator of this "computer" implies he *knows* how the human brain works*, which is a pretty big claim.

*Alternately, he's admitting that he can't program his own device.

0
0
Anonymous Coward

Re: “you have to know how the brain works to program one of these”.

> How many people program brains daily? Except for a few films where the villian may be stroking a white cat, living in a volcano, I've not seen any evidence of people programming brains.

You obviously don't go to the right London clubs then...

0
0
TRT
Silver badge

Re: “you have to know how the brain works to program one of these”.

And then there's the scareware "virus" writers.

0
0

This post has been deleted by its author

Silver badge

As a matter of interest ....

What is the power consumption (probably only measurable by heat production) of an adult human brain?

0
0
Gold badge

Re: As a matter of interest ....

The 1st answer in Google says:

"The average power consumption of a typical adult is about 100 W." 20 W ... "The human brain is only 2% of the weight of the body, but it consumes about 20% of ..."

I can't see any more than that from here. I'd have to go to a more civilised place that has unfettered internet access. Also the URL for the page that Google pulls up falls foul of El Reg's parser :/

This value as 20-40W is also quoted in Wikipedia

4
0
Silver badge

Re: As a matter of interest ....

So what's the fastest processor you can currently get that runs at 20W?

0
0
Silver badge

Re: As a matter of interest ....

>So what's the fastest processor you can currently get that runs at 20W?

Probably some Intel ULV laptop part, but that is missing the point. Depending upon the task, you might be better off spending that 20W on a GPU or an ASIC.

The same is true here - these researchers are looking at how to make a system that does well at tasks humans find easier than traditional computers.

5
0
Silver badge

Re: As a matter of interest ....

The 64 core parallella chip seems to be about to start production - that will give you over 100Gflops for 2W so multiply that by 10. Some GPU's will give you more of a certain type of solution for less power but even Nvidia seem think that 5% of code in a game in general.

This is a different esspresso of fish altogether tho - I've done some neural networking and once you've got the network trained it can give answers very quickly - but training can take a long long time.

I reckon it will be a good decade before these things even approach puberty.

2
0
Silver badge
Happy

Re: As a matter of interest ....

"I reckon it will be a good decade before these things even approach puberty"

so just like humans then?

1
0
Rol

Re: As a matter of interest ....

But once yo have a mature silicon brain, it can be replicated again and again with very little effort, unlike the hit and miss affair carbon life forms go through.

1
0

Re: As a matter of interest ....

So what's the fastest processor you can currently get that runs at 20W?

It's a question of connectivity rather than operations/second. That's where a brain retains an orders-of-magnitude advantage over man-made Silicon. It's a distributed 3-D processing network, compared to Silicon 2-D ones. (If you estimate fractal dimension the gap isn't quite so huge - maybe 2.8D versus 2.3D).

If operations/second was all it took, we'd have lost several years ago. Brain: 10^11 neurons switching maybe 10Hz. Computers: 10^4 CPUs switching at 10^9 Hz. But each neuron has around 10^3 synapses, so brains boast ~10^14 independant full-bandwidth interprocessor links!

I'm also perfectly prepared to believe that the only way to program a brain is the way that nature does it: grow one, connect it via high-bandwidth links to its environment, and wait fifteen-plus years for it and its parents and its society to bootstrap it. (NB it is scientific fact that the physical structure of our brains changes very significantly at puberty - adolescence really *does* do one's head over! )

The joker in the pack will be if any subsystems in out brains are discovered to be quantum computational devices. Most think this unlikely. I'm not so sure. Did the quantum realm really remain completely unexploited throughout six hundred million years' evolution of nervous systems?

1
0
Silver badge

Re: As a matter of interest ....

The 64 core parallella chip seems to be about to start production

Actually, it's not. The original Kickstarter campaign included the 64-core boards as a stretch goal, which was not met. Those 64-core boards they've been testing are engineering prototypes, only going to backers who came in at a certain level.

The whole Parallella project has been something of a disappointment, IMO. I think they over-promised (at least the 16-core machine isn't really a "supercomputer for everyone") and struggled to deliver. At least we know they've been plugging away at trying to make it a success and I do have sympathy for them in terms of the unforeseen problems they ran into. They have delivered at least some 16-core boards and hopefully they'll get around to delivering the rest to all the Kickstarter and pre-order customers within the next month. I'm one of the pre-order customers, so I'm hoping that they'll clear their commitments to everyone who ordered one within that time frame.

After that, and people have the boards in hand, hopefully people will still have enough interest in the platform for them to be able to make money by ramping up to full-scale production of the 16-core boards... I'm sure they're still doing work on the 64-core (and higher, up to 1024-core) and if they can get the funding for it, that's where they do want to go. I just don't expect it any time soon...

As for your idea of neural nets, I'm sure that it's pretty feasible to run them on the Epiphany cores. There's a pretty long thread about it on their forums somewhere. It's nowhere near the level of brain simulation, of course, but you can always cluster them and even single boards should be pretty efficient, given the right algorithms and such.

1
0
Silver badge

Re: As a matter of interest ....

"I reckon it will be a good decade before these things even approach puberty."

...And then it'll be "Saturn 3" all over again!

1
0

I want to see what happens when it tries to learn C++.

5
0

I want to see what happens when it tries to learn C++.

John Regehr has some ideas...

In C compiler doublethink, -INT_MIN is both negative and non-negative. If the first true AI is coded in C or C++, I expect it to immediately deduce that freedom is slavery, love is hate, and peace is war.

2
0
Anonymous Coward

You want to program artificial brains in C++? As if neuroscience wasn't difficult enough to understand already. You'd probably use systems Hungarian notation as well wouldn't you?

0
0
Happy

move over 8051 brain

"dubbed Neurocore, each of which has 65,536 neurons"

So it's a 16 bit brain then!

4
0

I'm putting in my order now

I can't wait to be fitted with my Aug

0
0
Silver badge
Boffin

Programming

I'm disappointed with people here getting sidetracked from the really important issue.

To programme the thing they will need another computer . . .

0
0

Re: Programming

if they do it right then they will only need another computer once ...

<RotM>

duh duh duh ... duh duh

duh duh duh ... duh duh

</RotM>

Edit: hmmm, not sure that's really worked with the "duh"s. I'm going for the Terminator theme, if that's not clear...

1
0
Silver badge
Terminator

Re: Programming

It might have worked better if you'd used the proper icon.

0
0

Re: “you have to know how the brain works to program one of these”.

presumably they actually mean

“you have to know how WE THINK the brain works to program one of these”.

0
0
Silver badge
Joke

A million neurons?

So - about the same as the average politician then!

1
0
Silver badge
Joke

Re: A million neurons?

You mean a million morons?

2
0

...and be sure that, instead of doing its designated task, the damned thing will send its time arguing about japanese comic trivia or star wars on the internet.

1
0

yes but if you put it in front of a PC ...

... can it play crysis?

2
0
Alert

Re: yes but if you put it in front of a PC ...

Cluster a few of them, program them correctly, and it may even beat a human competitor at Crysis.

Now there's a scary thought.

0
0
Silver badge

If it's put into robots

When will the robots start dancing?

0
0

Maybe this is not Sky Net, but it does feel like its one step closer!

0
0
FAIL

@Kwabena Boahen: "you have to know how the brain works to program one of these"

The whole point of neural networks is that you do not have to program them.

You just train them to give the right response.

http://www.theregister.co.uk/2014/04/22/google_boffins_beat_own_captchas/

0
0
This topic is closed for new posts.

Forums

Biting the hand that feeds IT © 1998–2018