Scientists model a problem-solving brain
The unstoppable rise in computing power has given scientists a novel method of investigating the human body. The computing industry has kept itself in accordance with Moore’s Law, doubling the speed of processors every two years. Taking advantage of the increasing computational power, scientists have been modelling the brain at the neuronal level in an attempt to under how it works. Despite much media hype of promises of animal-like behavior, past models have failed to come anywhere close to a real brain.
A group of scientists at the University of Waterloo (Ontario, Canada), however, are proving that there is more to modeling the brain in order to understand it than simply putting together some virtual neurons. Using a new computer called SPAUN, scientists have created a “brain,” containing about 2.5 million virtual neurons, capable of doing some basic functions nearly as well as humans. On the other hand, previous modeling attempts used upwards of 1 billion neurons with worse results.
Modeling neurons is a tricky task – computers cannot be hardwired to simulate a neuron. The two operates by completely different mechanisms. Whereas computers communicate information using ones and zeroes, neurons signal each other using action potentials, or sudden but fleeting changes in the cells’ voltage. Information is encoded in the frequency of the action potentials. In computers, neurons must be modeled in code and then be put together.
The key to SPAUN is not sheer processing power, but architecture. Just as how the architecture of a computer processor can impact its processing speed given a constant clock rate, the architecture of the virtual neurons have a direct impact on the capabilities of the virtual brain. In SPAUN, the virtual neurons are separated into functional groups. One group of neurons may be responsible for interpreting visual stimuli, another for memory, and yet another for output.
In their experiments, the scientists equipped SPAUN with a simple camera, which allowed it to receive visual instructions, and a robot arm, which allowed it to physically draw out its responses. SPAUN was instructed to perform a variety of tasks, such as recognizing letters and making inferences. The inference-task gave SPAUN two sequences of numbers. Using the given sequences, SPAUN predicted a missing number in a third sequence.
After the visual group processed the incoming command, another group of neurons determined which task was asked of the computer. The remaining neurons were then “rewired” to handle the task. A memory group, for example, retrieved the information applicable to the task. The “motor cortex” then commanded the robot arm to draw out SPAUN’s response. SPAUN came rather close to humans in performance – in character recognition, SPAUN was about 94% correct, whereas human were right about 98% of the time.
Despite the flexibility of SPAUN, the scientists admit that it is still very limited. Although SPAUN is able to determine what to do from a given input, it cannot learn any new tasks. All eight of the tasks SPAUN could perform are pre-coded into SPAUN’s memory.
However, SPAUN is only in its infancy. The 2.5 million neurons are paltry compared to the over 1 billion neurons used by the Blue Gene supercomputer to model a cat brain. Given more neurons, SPAUN would most definitely be able to learn more tasks. Perhaps soon, we may finally be able to model a working and thinking brain in a computer.