2
28 | NewScientist | 8 November 2014 What does it mean to build a computer that works like the human brain? We are trying to approximate the brain’s essence within today’s silicon technology. We’re not looking to model the human brain anatomically, but to build a computer that mimics its abilities for sensation, perception, interaction and cognition while rivalling its low power consumption and compact size. Current computers are very logical, sequential and quantitative. They are based on a 70-year-old architecture that separates memory from processing, and works in a step-by-step fashion, executing a series of pre-written “if X then do Y” equations. They are fast number crunchers and can process lots of data, but they really don’t think. That’s what we are trying to change. At the heart of this is a new kind of chip. How does it work? Conventional computers consume a lot of energy moving data to and from memory. But in the brain, memory and computation are not separated; they are closely intertwined. That’s the blueprint we used for the TrueNorth chip. We call it that because, like true north on a map, we believe it’s a clear, reliable reference and points the way to low- power systems that function like the brain. The brain’s cortex is made of microcolumns of between 100 and 200 neurons tiled together. This structure inspired the latest version of the chip, which is a network of more than 4000 cores each made up of 256 “neurons” connected by more than 65,000 “synapses”. Like our network of neurons in the brain, these neurosynaptic cores integrate memory, computation and communication. And they are also adaptable. How are they adaptable? The chips within modern computers can each have more than a billion transistors. If one transistor burns out in the computer on your desk, the chip stops working. But if one of the cores on the TrueNorth chip burns out, information is just routed around it. It’s similar to the way the brain can reroute around an injury. All this uses less power than current computers? This is the largest chip IBM has ever built – with 5.4 billion transistors – yet it consumes just 70 milliwatts of power when computing. A typical chip, like the one in your desktop computer, consumes about 70 watts – so 1000 times more power. The TrueNorth chip uses less power because it works like the brain; it performs operations only when it needs to and does many things in parallel. The idea to build a computer that mimics the brain has been around for decades. Why is it only truly starting to take off now? There is a push and a pull. The pull is that we have covered the planet with sensors, cameras, microphones, you name it. All that data is coming at us fast and furious. But with today’s computers, it’s difficult to analyse the data, make sense of it and act on it in real time. The push comes from our technological readiness. We are far from fathoming the true depths of how the brain works. But after studying it for more than 125 years, we have enough glimmers of insight into anatomy, physiology and behaviour that efforts to approximate these are now meaningful. With supercomputers, we are now able to carry out a simulation of a human brain. So we can study models of the brain in terms of computation, communication, speed, area, power and memory. And finally, transistors have now shrunk down to a point where we can attempt radically new architectures. Why do we need a new architecture when the old one – used on supercomputers like IBM’s Watson – is still yielding impressive results? Cognitive computers, like Watson, combine artificial intelligence and machine-learning algorithms, which get better the more data you feed into them. The strategy with these has been to take the brain’s behaviour and try to reproduce it. But complex problems of seeing and understanding – recognising your mother’s PrOfiLe Dharmendra Modha is the founder of the Cognitive Computing group at IBM Research- Almaden in San Jose, California, and the principal investigator for the US Department of Defense’s project aiming to build scalable neurocomputers OPINION INTERVIEW A computer that thinks Computers are great for cold calculations, but we are still far better at interpreting nuance. IBM researcher Dharmendra Modha says the brain is a blueprint for a better solution

A computer that thinks

  • Upload
    clint

  • View
    215

  • Download
    1

Embed Size (px)

Citation preview

Page 1: A computer that thinks

28 | NewScientist | 8 November 2014

What does it mean to build a computer that works like the human brain?We are trying to approximate the brain’s essence within today’s silicon technology. We’re not looking to model the human brain anatomically, but to build a computer that mimics its abilities for sensation, perception, interaction and cognition while rivalling its low power consumption and compact size.

Current computers are very logical, sequential and quantitative. They are based on a 70-year-old architecture that separates memory from processing, and works in a step-by-step fashion, executing a series of pre-written “if X then do Y” equations. They are fast number crunchers and can process lots of data, but they really don’t think. That’s what we are trying to change.

At the heart of this is a new kind of chip. How does it work?Conventional computers consume a lot of energy moving data to and from memory. But in the brain, memory and computation are not separated; they are closely intertwined. That’s the blueprint we used for the TrueNorth chip. We call it that because, like true north on a map, we believe it’s a clear, reliable reference and points the way to low-power systems that function like the brain.

The brain’s cortex is made of microcolumns of between 100 and 200 neurons tiled together. This structure inspired the latest version of the chip, which is a network of more than 4000 cores each made up of 256 “neurons” connected by more than 65,000 “synapses”. Like our network of neurons in the brain, these neurosynaptic cores integrate memory, computation and communication. And they are also adaptable.

How are they adaptable?The chips within modern computers can each have more than a billion transistors. If one transistor burns out in the computer on your desk, the chip stops working. But if one of the cores on the TrueNorth chip burns out, information is just routed around it. It’s similar to the way the brain can reroute around an injury.

All this uses less power than current computers?This is the largest chip IBM has ever built – with 5.4 billion transistors – yet it consumes just 70 milliwatts of power when computing. A typical chip, like the one in your desktop computer, consumes about 70 watts – so 1000 times more power. The TrueNorth chip uses less power because it works like the brain; it performs operations only when it needs to and does many things in parallel.

The idea to build a computer that mimics the brain has been around for decades. Why is it only truly starting to take off now?There is a push and a pull. The pull is that we have covered the planet with sensors, cameras, microphones, you name it. All that data is coming at us fast and furious. But with today’s computers, it’s difficult to analyse the data, make sense of it and act on it in real time.

The push comes from our technological readiness. We are far from fathoming the true depths of how the brain works. But after studying it for more than 125 years, we have enough glimmers of insight into anatomy, physiology and behaviour that efforts to approximate these are now meaningful. With supercomputers, we are now able to carry out a simulation of a human brain. So we can study models of the brain in terms of

computation, communication, speed, area, power and memory. And finally, transistors have now shrunk down to a point where we can attempt radically new architectures.

Why do we need a new architecture when the old one – used on supercomputers like IBM’s Watson – is still yielding impressive results?Cognitive computers, like Watson, combine artificial intelligence and machine-learning algorithms, which get better the more data you feed into them. The strategy with these has been to take the brain’s behaviour and try to reproduce it.

But complex problems of seeing and understanding – recognising your mother’s

ProfileDharmendra Modha is the founder of the Cognitive Computing group at IBM Research-Almaden in San Jose, California, and the principal investigator for the US Department of Defense’s project aiming to build scalable neurocomputers

OPINION INTERVIEW

A computer that thinksComputers are great for cold calculations, but we are still far better at interpreting nuance. IBM researcher Dharmendra Modha says the brain is a blueprint for a better solution

141108_Op_Interview.indd 28 31/10/2014 16:00

Page 2: A computer that thinks

8 November 2014 | NewScientist | 29

face in a crowd despite the fact that she’s changed her hairstyle or the light that day is different, for instance – are impossible for today’s computers. For these types of problems, we need new programs that run on different types of algorithms. And to invent these we need to rethink the underlying architecture.

Is computing power also an issue?Yes. Practically, we need a new architecture because brain-like computation is extremely inefficient at the moment. With the most powerful supercomputers we can run simulations of the brain 1500 times slower than real time. On a standard computer,

a real-time simulation at the scale of 100 trillion synapses – which is how many you would find in the average human brain – would require about 12 gigawatts of power. That’s more than the power consumption of London and Los Angeles put together. Our brains can carry out the same computation at 20 watts – like powering a dim light bulb. Extracting brain-like functionality from today’s computers within reasonable constraints of power, volume, speed and cost is just not possible.

So would this new computer architecture replace existing designs altogether?No. These are two complementary elements of

computing that we need as a society. Think of them like yin and yang. For certain functions we will continue to rely on the brute strength of traditional computers; for other tasks we will want computers that can think and adapt.

How can you design a computer based on the human brain when our understanding of it is still so far from complete?

When the accomplished physicist Richard Feynman died, on his whiteboard

was written: “What I cannot build, I cannot understand.” It is

impossible to record all the activity of the brain in a live, normally behaving animal, so it will give up its secrets slowly. But taking what we do know and

casting it into a computational model has two benefits. It gives you a tool to test hypotheses for how the brain might work, and it gives you a new technology that, although just a cartoon of the brain, is already a powerful machine. Also, by starting to build these types of computers, we speed up progress in computation and neuroscience.

Why do we need computers that work like our brains when there are 7 billion brains on the planet, many of them underemployed?How many volunteers are there who are willing to walk into the Fukushima plant right now? When faced with disasters, such as a burning house or chemical spill, we don’t want to put humans at risk. Or imagine creating a traffic camera that watches cars and can detect whether an accident is about to happen just by looking at the conflicting movements and velocities. How many people would be able to hold their attention flawlessly on hundreds of millions of traffic cameras?

The sensory information that is flowing at us now is so vast that the task of pattern recognition, while amazingly useful, is in the end repetitive and has to be given to machines.

How do you see brain-inspired computers shaping the world in years to come?All futurists are destined to either fail or predict what they already know is coming true. So rather than predicting the future, I’d like to focus my attention on creating it. n

Interview by Clint Witchalls

Ibm

“ To simulate the human brain would use more power than London and Los Angeles”

For more opinion articles, visit newscientist.com/opinion

141108_Op_Interview.indd 29 31/10/2014 17:50