To err should be inhuman

摘要

Neural networks, like the ones grabbing headlines for winning boardgames or driving cars, depend on huge amounts of computing hardware. That in turn means a colossal amount of power: the next wave may consume millions of watts each. That's one reason why some suggest we rethink what we want computers to be. Reducing the precision with which they analyse problems, and putting up with the odd "error", can cut zeroes off their energy consumption (see page 36). And it has precedent in the human brain - an unrivalled piece of hardware using electrical fluctuations and requiring a million times less power than a computer.