How SPAUN Works

Jonathan Strickland

© Carson Ganci/Design Pics/Corbis
© Carson Ganci/Design Pics/Corbis

If there's one thing all those Frankenstein movies have taught me, it's that the brain is a tricky thing. It's so tricky that even today with all our computing power and knowledge we can't create a full, working simulation of the human brain. That hasn't stopped computer scientists and neurologists from giving it the old college try.

One group that has shown incredible progress in simulating parts of a brain is a research team at the University of Waterloo. They've built a simulation they call the Semantic Pointer Architecture Unified Network, or SPAUN for short. SPAUN is a computer model of a human brain. It has 2.5 million virtual neurons -- a typical human brain has 100 billion neurons.

The researchers organized the neurons into structures that mimic parts of the human brain. Specifically, they built models of the prefrontal cortex, the thalamus and the basal ganglia. Collectively, these simulated elements of a human brain can detect, process and respond to information. The virtual brain responds by manipulating a physically-modeled arm.

What do these parts of the brain do? The prefrontal cortex analyzes thoughts and conducts abstract thinking. It's the home of our working memory and we use it for decision-making processes. The thalamus relays perceptual information and governs motor control. The basal ganglia are important for learning and motor control.

What this means is that SPAUN can process visual information -- such as a string of numbers -- and act on that information in some way. For example, you could present SPAUN with a series of numbers and ask it to identify the third figure in the series. SPAUN would then direct the mechanical arm to draw the appropriate figure.

The virtual brain is also able to recognize different forms of the same figure. The way you write the number 2 and the way I do may not look the same but we would both recognize the figure as being a 2. This is tricky for computers, which tend to be able to recognize shapes only if they fall within a fairly narrow set of parameters. But SPAUN can analyze the shape and match it to the corresponding figure.

What I think is the most interesting aspect of SPAUN is how it makes mistakes. Just like humans, if you give SPAUN a long enough series of numbers it starts to have trouble remembering them all. And also like humans, SPAUN tends to be better at remembering the first few and last few numbers in a series -- the ones in the middle give it trouble.

The model is still relatively simple, particularly compared to a human brain. And it's not easy on the computing power either -- every second of virtual brain activity requires two hours of processing on a supercomputer. But the researchers hope to harness a supercomputer better aligned to the functions of SPAUN with a goal of reaching a one-to-one ratio of brain activity to processing time.

Virtual models like SPAUN help in multiple disciplines. Not only are we learning more about artificial intelligence, but also aspects of our own intelligence. As we explore these areas further we might discover new ways to build smarter machines. Or maybe we'll build machines that seem human to us because they possess the same limitations and make similar mistakes. Perhaps we'll address our own limitations through a greater understanding of how our brains work!

Here's a quick video of how SPAUN processes and acts on information.