New WBE implementation

post by Louie · 2012-11-30T11:16:27.096Z · LW · GW · Legacy · 7 comments

Contents

7 comments

It usually isn't profitable to pay attention to science news, since science journalists largely misintrepret new "breakthroughs". But I am somewhat interested in this story about "artificial brains" coming out of Canada.

Most large neuron simulations I've read about before don't actually do anything. But apparently there's a somewhat large new WBE implementation at the University of Waterloo that performs sub-humanly on several tasks while having similar weaknesses to human brains.

Curious what others think of this recent development.

 

7 comments

Comments sorted by top scores.

comment by AlexMennen · 2012-11-30T17:46:47.196Z · LW(p) · GW(p)

This is neuromorphic AI, not WBE. Still, this is interesting and terrifying.

comment by jimrandomh · 2012-11-30T15:56:59.020Z · LW(p) · GW(p)

The actual project, links to including source code, is here. A link to the paper this article is based on, which I found on HN, here.

Replies from: Wei_Dai
↑ comment by Wei Dai (Wei_Dai) · 2012-12-02T18:04:15.907Z · LW(p) · GW(p)

I found a more detailed paper here which actually explains how Spaun works. What they're doing is really pretty far from WBE or even neuromorphic AI. They're wrote down specific algorithms for solving the tasks they wanted Spaun to perform, and then analytically solved for neuron connection weights that would cause their model to implement those algorithms, so essentially just using the simulated brain as a very slow programmable computer.

comment by Wei Dai (Wei_Dai) · 2012-12-02T18:48:50.362Z · LW(p) · GW(p)

This isn't WBE or even AGI, because they essentially programed the simulated brain to perform certain specific tasks, by writing down algorithms for doing those tasks and then choosing synaptic connection weights to implement those algorithms. (See this paper for details.) Neuromorphic AI is scary because someone might be able to build an AGI without having to understand general intelligence on an algorithmic level, but that's not how Spaun works, at least not the current iteration of it.

comment by Kingoftheinternet · 2012-11-30T16:00:46.906Z · LW(p) · GW(p)

Here's a collection of videos by the researchers on what exactly this thing does. I'm impressed, excited, and worried all at once.

comment by Risto_Saarelma · 2012-12-03T18:59:48.887Z · LW(p) · GW(p)

The researchers are currently doing a Reddit AmA.

comment by Dr_Manhattan · 2012-11-30T13:50:57.309Z · LW(p) · GW(p)

Same as you - filed under the "interesting" pile. I think with results like this (published in Science according to Nature) we'll know much more soon as people are going to be lining up to throw more hardware at this.