SINCE nobody really knows how brains work, those researching them must often resort to analogies. A common one is that a brain is a sort of squishy, imprecise, biological version of a digital computer. But analogies work both ways, and computer scientists have a long history of trying to improve their creations by taking ideas from biology. The trendy and rapidly developing branch of artificial intelligence known as “deep learning”, for instance, takes much of its inspiration from the way biological brains are put together.
The general idea of building computers to resemble brains is called neuromorphic computing, a term coined by Carver Mead, a pioneering computer scientist, in the late 1980s. There are many attractions. Brains may be slow and error-prone, but they are also robust, adaptable and frugal. They excel at processing the sort of noisy, uncertain data that are common in the real world but which tend to give conventional electronic computers, with their prescriptive arithmetical approach, indigestion. The latest development in this area came on August 3rd, when a group of researchers led by Evangelos Eleftheriou at IBM’s research laboratory in Zurich announced, in a paper published in Nature Nanotechnology, that they had built a working, artificial version of a neuron.
Neurons are the spindly, highly interconnected cells that do most of the heavy lifting in real brains. The idea of making artificial versions of them is not new. Dr Mead himself has experimented with using specially tuned transistors, the tiny electronic switches that form the basis of computers, to mimic some of their behaviour. These days, though, the sorts of artificial neurons that do everything from serving advertisements on web pages to recognising faces in Facebook posts are mostly simulated in software, with the underlying code running on ordinary silicon. That works, but as any computer scientist will tell you, creating an ersatz version of something in software is inevitably less precise and more computationally costly than simply making use of the thing itself.
Hearing the noise, seeing the signal
Neurons are pattern-recognition devices. An individual neuron can be connected to dozens or hundreds of others, and can pass electrical signals to and fro. If it receives a sufficient number of strong enough signals from its brethren over a short enough span of time, it will “fire”, sending a jolt of electricity to other neurons connected to it, possibly causing them to fire as well. If the incoming signals are too weak, or too infrequent, it will remain quiescent.
Dr Eleftheriou’s invention consists of a tiny blob of germanium antimony telluride sandwiched between two electrodes. Germanium antimony telluride is what is known as a phase-change material. This means that its physical structure alters as electricity passes through it. It starts off as a disordered blob that lacks any regular atomic structure, and which conducts electricity poorly. If a low-voltage electrical jolt is applied, though, a small portion of the stuff will heat up and rearrange itself into an ordered crystal with much higher conductivity. Apply enough such jolts and most of the blob will become conductive, at which point current can pass through it and the neuron fires, just like the real thing. A high-voltage current can then be applied to melt the crystals back down and reset the neuron.
This arrangement mimics real neurons in another way, too. Neurons are unpredictable. Fluctuations within the cell mean a given input will not always produce the same output. To an electronic engineer, that is anathema. But, says Tomas Tuma, the paper’s lead author, nature makes clever use of this randomness to let groups of neurons accomplish things that they could not if they were perfectly predictable. They can, for instance, jiggle a system out of a mathematical trap called a local minimum where a digital computer’s algorithms might get stuck. Software neurons must have their randomness injected artificially. But since the precise atomic details of the crystallisation process in IBM’s ersatz neurons differ from cycle to cycle, their behaviour is necessarily slightly unpredictable.
The team have put their electronic neurons through their paces. A single artificial neuron, hooked up to the appropriate inputs, was able, reliably, to identify patterns in noisy, jittery test data. Dr Tuma is confident that, with modern chip-making techniques, his neurons can be made far smaller than the equivalent amount of conventional circuitry—and that they should consume much less power.
The next step, says Dr Eleftheriou, is to experiment with linking such neurons into networks. Small versions of these networks could be attached to sensors and tuned to detect anything from, say, unusual temperatures in factory machinery, to worrying electrical rhythms in a patient’s heart, to specific types of trade in financial markets. Bigger versions could be baked onto standard computer chips, offering a fast, frugal co-processor designed to excel at pattern-recognition tasks—like speech- or face-recognition—now performed by slower, less efficient software running on standard circuitry. Do that and the conceptual gap between artificial brains and real ones will shrink a little further.
Original Article By Economist