Electronic devices that mimic how brain cells in a cat work could allow computers to one day learn and recognize information more like humans do.
Such brain-like devices might accomplish more complex decisions and perform more tasks simultaneously than conventional computers are capable of, researchers added.
"We are building a computer in the same way that nature builds a brain," said researcher Wei Lu, a computer engineer at the University of Michigan.
Microchips typically rely on transistors, which are essentially switches that can flick on or off to represent data as the binary digits or bits 0 and 1.
The devices that investigators at the University of Michigan are developing instead employ "memristors." These circuit elements, unlike others, carry memories of their past: When you turn off voltage to the device, memristors remember how much was applied beforehand and for how long.
The very nature of memristors makes them act very much like synapses, which connect brain cells, or neurons, together. Synapses serve as reconfigurable switches that can form pathways linking thousands of neurons, and like memristors, they remember these pathways based on the strength and timing of electrical signals they receive from the neurons.
One of the world's most sophisticated supercomputers — Dawn, at Lawrence Livermore National Laboratory — can simulate 1 billion neurons and 10 trillion synapses, exceeding the scale of a cat brain. Still, it is a massive machine with more than 140,000 central processing units that needs a million watts of electricity and it still performs 100 to 1,000 times slower than a cat's brain.
In a conventional computer, the elements for logic and memory are located in different parts of the circuit and each element is only connected to a handful of neighbors in the circuit. As a result, conventional computers work in a linear fashion, line by line, making them excellent at performing relatively simple tasks with limited variables.
A brain, on the other hand can perform many operations simultaneously, or in parallel. This enables us to recognize a face in an instant, but even a supercomputer would take far longer and consume much more energy in trying.
Now Lu has used memristors to link conventional circuits together to mimic the brain.
"The idea is to use a completely different paradigm compared to conventional computers," he explained.
The system Lu and his colleagues devised, which connects two electronic circuits with one memristor — mimicking two neurons and a synapse — is capable of a memory and learning process with the esoteric name of "spike timing dependent plasticity." This refers to the ability of connections between neurons to become stronger when they are stimulated in relation to each other, and is thought to be the basis for memory and learning in mammalian brains.
Specifically, the researchers showed they could vary the duration and sequence of voltage they applied to their system to gradually increase or decrease the level of its electrical conductivity.
"In our brains, similar changes in synapse conductance essentially give rise to long term memory," Lu said.
The scientists are aiming toward an electric brain as smart as a cat — for instance, one that can figure out the shortest route from the front door to the sofa in a house full of furniture time after time, even if one moved the sofa each time.
"I can imagine the Department of Defense may be interested in putting such intelligent computers on weapons or unmanned vehicles such as drones so that they can make decisions directly without having to transmitting images back to controllers and waiting for commands," Lu said.
The next step is to build a larger system with hundreds of artificial neurons and memristor synapses, Lu said. The ultimate goal would be achieving the sophistication of a supercomputer in a machine the size of a two-liter bottle of soda.
"It is now possible to build a brain-like computer using electronic components — namely, transistors and memristors," Lu told TechNewsDaily.
Not just supercomputers will benefit
Aside from developing computers that behave more like brains, memristors could also help conventional computers continue to keep pace with Moore's Law, which states that computing power should double every two years on average.
In research appearing in April 8 issue of the journal Nature, researchers at Hewlett-Packard Laboratories showed that memristors could perform computations, meaning that logic functions could take place in chips where data is stored, instead of being separated away on a specialized central processing unit, as is the current practice.
In other words, future increases in computing power might not come from an increase in raw processor speed, but "from the increase in computing efficiency instead," Lu said.
Lu and his colleagues will detail their findings in the April issue of the journal Nano Letters.