Brain-like chip outstrips normal computers

COMPUTER chips that mimic the human brain are outstripping conventional chips in crucial ways. They could also revolutionise our understanding of how the brain functions.

Attempts to simulate the brain usually involve programming software to behave like groups of neurons. A new "neuromorphic" design instead tries to recreate the brain's hardware, using analogue components last seen in the early days of computing. "On our system, you can physically point to the neuron," says Karlheinz Meier of the University of Heidelberg in Germany.

The Spikey chip contains 400 "neurons", or printed circuits. Real neurons have a voltage across their outer membrane, which Spikey mimics using capacitors: components that store charge. Just as in a real neuron, when the applied voltage reaches a certain level, the capacitor becomes conductive, firing a "nerve signal".

Spikey also mimics synapses - the connections between neurons. In a normal chip, every process is digital and so can only take the value 0 or 1. Meier's team instead used analogue components with variable levels of resistance to simulate the way connections between neurons become stronger or weaker depending on how much they are used. "Analogue circuits died after digital computers became more powerful," says Meier, but they are now finding new roles.

The team connected the neurons in the Spikey chip in different ways to mimic various brain circuits. They have now modelled six neural networks, including one found in the insect olfactory system. By measuring patterns of activity, they found Spikey's artificial networks behave much like the real thing (arxiv.org/abs/1210.7083).

"This is as good as you can get in simulating neural architecture," says Massimiliano Versace of Boston University, Massachusetts.

Neuromorphic chips do already exist, though until now each chip could only mimic one particular brain circuit. Spikey, on the other hand, can recreate any pattern.

Neuromorphics have advantages over conventional chips that makes them useful in certain situations. For example, they do not separate memory and computation - information is stored in the synaptic strength - so they can run faster using less power. They also cope better with damage. Knocking out a few bits of a normal chip often breaks it altogether, but neuromorphics keep working, albeit slowly.

Companies like IBM and HP are looking into neuromorphics (see "Compute like a human"), and some medical devices already use them. Versace is working with NASA to develop a neuromorphic system to control a Mars rover, and says that the chips' fault tolerance may make them better suited to surviving the intense radiation of space. The chips also allow theories of how the brain functions to be tested, in experiments that systematically change how each neuron and network behaves.

The team is now scaling up Spikey as part of a project called BrainScales. "Instead of 400 neurons we have 200,000," says team member Thomas Pfeil. The researchers have printed all the circuits onto a single silicon wafer, 20 centimetres across, which allowed them to incorporate many more connections. Next year, they will use it to simulate part of the cortex of a rat brain. From there, they plan to connect six wafers in parallel, simulating over a million neurons, and eventually model a rat's entire visual cortex.