THE THINKING ROBOT: YOUR NEW BEST FRIEND?
Scientists at the University of Waterloo, Canada, led by Professor Chris Eliasmith, have built the most sophisticated simula
tion of a working brain ever constructed. Although much smaller than the human brain itself, consisting of only 2.5 million brain cells (compared to 100 billion) and many fewer than some previous simula
tions – it displays an impressive range of different behaviours. The artificial brain can recognise images, remember sequences, and even complete the kind of complex task you might find in an IQ test. The supercomputer program, called SPAUN (Semantic Pointer Architecture Unified Network), uses a 28 x 28 pixel digital camera ‘eye’ to gather input from its surroundings and then gives its responses with a robot arm. For instance, when shown the sequence 1 2 3 - 5 6 7 - 3 4 ?, together with an instruction, SPAUN will scrawl the digit ‘5’ on a piece of paper. Unlike IBM’s Watson supercomputer, which was built in 2011 to do one thing (play Jeopardy!) and do it well, but made
no attempt to copy how the brain works, SPAUN replicates actual brain cell activity and wiring. More importantly, though, it turns this activity into behaviour. This is in contrast to larger, more detailed brain models, such as the Blue Brain Project, which produce detailed simulations of neural activity, but don’t necessarily do anything. SPAUN has two software systems that
work in harmony: a ‘working memory’ system that is modelled on the ‘higher’ thinking part of the brain (called the prefrontal cortex – where we make our decisions), and an ‘action selection’ system, which is based on other parts of the brain called the basal ganglia and thalamus (more primitive, instinctual regions). The ‘action
selection’ system routes data to the part of the ‘working memory’ system appropriate for the task, which then stores that data, and performs the necessary ‘thinking’ functions. All this is accomplished by software running on a supercomputer, but, as it stands, is still very limited compared to a real brain: SPAUN can only tackle eight predefined tasks, and is far slower than the real thing, taking around two hours of computing time to simulate one second of brain activity. It does however demonstrate a range of cognitive skills and even makes some of the same mistakes we do, such as remembering the first and last items in a list better than the others (known to psychologists as primacy and recency). The team also looked at the effects of ‘killing off’ neurons to simulate ageing, and have seen patterns of decline similar to what happens in old age. Crucially though, SPAUN lacks adaptivity – the ability to learn new tasks. This is a shortcoming the team hopes to tackle in
future. Even so, there’s no reason to presume that building on this simulation will at some point produce the more elusive qualities of living brains, such as awareness or intentions: SPAUN doesn’t do any of its impressive tricks because it wants to – it is explicitly programmed and fed instructions, like any other computer system. So without free will, there’s probably not much need to worry about a future version taking control of our missile defence systems just yet…