home
RSS
Brain model may help build human-like robot
Can a computer mimic the way the brain works?
November 29th, 2012
02:02 PM ET

Brain model may help build human-like robot

It goes without saying that the human brain is complex, and would be hard to build from scratch. But researchers are looking to simulate how the brain works so that more human-like artificial intelligence can be created and we can better understand damage to our own brains.

Chris Eliasmith of the University of Waterloo in Ontario, Canada, led research published in the journal Science on a brain model called SPAUN - the Semantic Pointer Architecture Unified Network.

SPAUN lives inside a computer, can view images with a camera-like eye and can draw responses to questions. For example, show it the number "4" and it will write its own "4." It can even mimic the style of the numeral.

Both in the brain and in SPAUN, neurons communicate by changing their voltages, and the pattern of these voltage "spikes" is what carries information from one cell to another, Eliasmith said. The receiving cell generates a voltage of its own if it receives a particular voltage.

SPAUN has 2.5 million spiking neurons. Neurons are the cells - the individual components - that make up the brain. The human brain has about 100 billion neurons, so there's still a long way to go in terms of replicating its full capacity.

It's hard to compare SPAUN to any existing animal. Monkeys can do more general recognition than what this model does, Eliasmith said. But there are tasks SPAUN does that, until now, it was thought only humans could do.

"It's not as smart as monkeys when it comes to categorization, but it's actually smarter than monkeys when it comes to recognizing syntactic patterns, structured patterns in the input, that monkeys won't recognize," Eliasmith said.

All of SPAUN's tasks involve numbers. For instance, give it the pattern: 1, 11, 111; 3, 33, 333; 4, 44, _____." SPAUN could fill in the blank as 444.

"That's actually part of an intelligence test, realizing that everything is increasing by one," Eliasmith said. "Monkeys actually won't figure that out."

A drawback is that SPAUN cannot operate in real time. For every one second in an online demonstration video, it takes 2.5 hours. The researchers are hoping to be able to get it to do real-time operations. And SPAUN, as a simulation of actual neurons, is not a robot.

But structuring the artificially intelligent brain like a human brain means that the kinds of errors it makes are the kind of that people make, and its reaction time - how long it takes to "think" about problems - would be similar to humans. This could help with the creation of robots.

"All of that kind of thing will make for the possibility of having agents that are more human-like to interact with," Eliasmith said.

Christian Machens, neuroscientist at the Champalimaud Neuroscience Programme in Lisbon, Portugal, points out in an accompanying piece in Science that SPAUN is not able to learn any new tasks. Its knowledge is entirely hard-wired.

Still, Machens writes that the authors offer a "coherent theory" of the workings of the brain (except for learning). And it sets a new goal for simulations: "to not simply incorporate the largest number of neurons or the greatest amount of detail, but to reproduce the largest amount of functionality and behavior."

This research is also useful for modeling brain damage, Eliasmith said. In separate research, he and colleagues looked at what happens when neurons in a simulation get destroyed at the same rate as in humans as they age.

"We can show that the performance on (an) intelligence task mimics the kind of performance that you see in people; it gets worse in the same sort of proportion that you find in humans," he said.

As for whether we'll have a fully cognizant, self-aware artificial intelligence in our lifetimes, Eliasmith isn't sure, but that is the sort of goal that he and his team are working toward. And they've come up with a novel way of approaching the problem.

"It’s hard to know at this point whether this approach is going to hit some wall that we haven’t seen yet, or actually be able to reach that holy grail, as it were, of artificial intelligence.”


soundoff (10 Responses)
  1. anon

    Does this mean anything? It looks like a pretty impressive feat, but it's always hard to tell. Will brains like this put us all out of a job in a couple of years? wow

    November 29, 2012 at 20:06 | Report abuse | Reply
  2. Abucadazzar

    They could also imitate rare, hard to treat brain diseases and see what happens to the neurons in the brain during the span of the " disease".

    November 30, 2012 at 01:26 | Report abuse | Reply
    • Jacob Ewing

      I doubt this will lead to a good simulation for human neurological problems, at it's not the actual biological elements themselves, but a simulation of their behaviour. That is, from what I see, this simulates the input, behaviour and output of the neurons, but not the cell tissue, the mechanical components and so on. I think rats and pigs will remain ~far~ more accurate test cases, regardless of how much this system improves.

      November 30, 2012 at 15:17 | Report abuse |
    • Trevor Bekolay

      Actually, we do simulate all the way down to the level of individual neurons, and so we can simulate many brain diseases. Our neuron model is phenomenological, so it may not provide a perfect analogy to the real world, but it should be able to capture many of the features of brain diseases. We could simulate, for example, the slowing of signal propagation in demyelination (as in MS), or the decrease in tonic dopamine levels in Parkinson's disease. Of course, in order to simulate those diseases, you would first need to create a good model of the non-pathological system; that's what we're working our way towards.

      Note: I am one of the co-authors of the paper.

      November 30, 2012 at 16:02 | Report abuse |
  3. Michael C Hall

    How well can this pattern recognition scale to other problems? Can it be trained on learning patterns that are "good" and ignoring patterns that are just noise?

    November 30, 2012 at 09:46 | Report abuse | Reply
    • Trevor Bekolay

      The only actual training via classical machine learning techniques is in the vision system. Certainly the vision system is sensitive to the type of input used for training - in this case, we only used handwritten digits from the MNIST database. Other than that, everything is programmed in by state – action associations; so, in the state that you see a "2", and you're in the copy drawing task, the action to take is to write a "2". To do a new task, you just have to make a new set of rules like that.

      November 30, 2012 at 16:05 | Report abuse |
  4. scanboy

    Anyone interested in this article should also read about the work of J.R. Anderson and his ACT-R cognitive architecture at Carnegie Mellon.

    December 1, 2012 at 09:55 | Report abuse | Reply
  5. saveyourfeet

    ahhh All diabetics need to read this blog: saveyourfeetDOTwordpressDOTcom

    December 3, 2012 at 18:27 | Report abuse | Reply
  6. ABB

    Thanks for the post! Your suggestions are very interesting and useful. Home based business is not an easy task. It requires proper planning and the skills to start a business.

    ABB

    December 4, 2012 at 01:56 | Report abuse | Reply
  7. Caitlyn Wang

    Nice Blog and good Thought, Thanks for the post! Your suggestions are very interesting and useful. Home based business is not an easy task. It requires proper planning and the skills to start a business. Happy New year.....!!!!

    January 1, 2013 at 16:03 | Report abuse | Reply

Post a comment


 

CNN welcomes a lively and courteous discussion as long as you follow the Rules of Conduct set forth in our Terms of Service. Comments are not pre-screened before they post. You agree that anything you post may be used, along with your name and profile picture, in accordance with our Privacy Policy and the license you have granted pursuant to our Terms of Service.

Advertisement
About this blog

Get a behind-the-scenes look at the latest stories from CNN Chief Medical Correspondent, Dr. Sanjay Gupta, Senior Medical Correspondent Elizabeth Cohen and the CNN Medical Unit producers. They'll share news and views on health and medical trends - info that will help you take better care of yourself and the people you love.