I think a computer architecture concept intended to reduce power consumption may also have a lot of potential for hardware-level artificial intelligence implementations.
Palem's team has designed a format for embedded chips that are capable of going with a hunch rather than performing painfully precise calculations. Early testing indicates that the chips can extend battery life and, as a bonus, may also enable mobile devices to run more complex applications, since an awful lot of time is currently wasted in churning out those exact calculations. ...A lot of computational power is used in AI to simulate probabalistic unreliability, but it would be pretty neat if we could get it built in on the hardware.
In conventional computer circuits, each single bit of information is represented by a 0 or a 1. This information is definite; a circuit is either on or off -- the answer is either yes or no.
But with Probabilistic Bits, or PBits, chips, a circuit can be "on" with a high degree of certainty, but not with 100 percent certainty, said Palem, whose team developed the PBits prototype.
"In the chips being built today, the hardware obeys the software instructions absolutely, even though the application software does not require such precision," said Palem. "So the simple idea that we have is to make the following connection: If probabilistic algorithms do not need the hardware to be reliable, then why invest a lot of money and time in making hardware reliable?"