Credit: IBM
IBM has created artificial neurons that work in a way much closer to a human brain than previous efforts. It’s a step closer to creating computers that learn in the same way as humans.
The project is the latest attempt to overcome the biggest weakness of computers: that in principle a processor can only do one thing at a time. That means it is far quicker and more reliable than a human at straightforward tasks, but is limited when it comes to weighing up multiple options simultaneously.
Most attempts at artificial intelligence aim to replicate the neural network, the way by which a huge series of electrical connections in the human brain can effectively act as a flowchart or decision tree. While there’s been some success in this approach, it’s tough to replicate the scale of a brain in a processor. As IBM’s Abu Sebastian told Newsweek, a brain has something in the region of 100 billion neurons, each with 10,000 synapses, and trying to fit that number of electrical connections in a chip is practically challenging to say the least.
The IBM approach is to avoid the usual technique of recreating the brain through a digital network and instead take a more accurate analog approach. They’ve done it by making the artificial neurons from phase change materials, similar to the way Blu-ray discs are created. Applying heat or electricity changes the material between a crystalline and amorphous state and in turn determines whether it’s an electrical insulator or conductor.
As well as being tiny, quick to respond, and using little energy, the big advantage of these artificial neurons is that they have a degree of randomness. The precise structure of the material when in the amorphous state varies each time it makes the switch, in turn changing the time it takes to transform to a crystalline state, begin passing through electricity and thus affect the adjoining neurons. This partially random process is a replication of the way neurons in a brain ‘fire’.
On the scale of an entire brain, this small element of randomness makes thought more efficient by allowing it to effectively explore random samples of a range of possibilities, helping narrow down the range of likely effective answers or solutions.
At the moment the IBM artificial network uses a mere 500 neurons, but the company believes the principle works and could be scaled. It says computers working this way would be more effective at spotting patterns and connections.