DARPA-Funded Chip Calculates With Probabilities, Not Hard Binary Logic
Most people with even the most fundamental knowledge of how computer chips work are familiar with binary logic — the...
Most people with even the most fundamental knowledge of how computer chips work are familiar with binary logic — the system of ones and zeros that enable modern computing to occur — in which an input always results in a solid result (either a one or a zero). Now, a Boston-based startup is rewiring the basic concept of computation with a probability processor that deals in chance rather than binary logic, creating a chip that could speed all kinds of processes from flash memory in smartphones to better decision-making software for machines.
Lyric Semiconductor’s chip accepts probabilities as inputs instead of ones and zeros, and the output is also a probability — the odds that the two input probabilities match up. Rather than the usual NAND gates characteristic of conventional transistor schemes, the chips employ what are known as Bayesian NAND gates, named for the statistician Thomas Bayes whose field of thought is the basis for the idea.
The ability to crunch probabilities rather than hard ones and zeros is actually far more applicable to many computing tasks than binary logic, because many of said tasks are already trying to crunch probabilistic math. Recommendation engines like those employed by Amazon and Netflix, for instance, rely on software figuring out what you are most likely to enjoy. Spam filters work in much the same way.
Of course, such processors might also have defense applications that are not lost on DARPA, which has been partially funding their development since 2006. Probability processors could be used to patch together distorted audio signals or to create better machine vision that can actually interpret what an image or action means. Ostensibly they could also be used to help intelligence services sort through vast amounts of data to separate the noise from valuable information.
Lyric plans to have prototypes of their all-purpose probability processors working within three years. In the meantime, a smaller flash memory error-correcting chip based on the technology is available for license this week, and the company hopes to have the chips boosting flash memory in portable devices like smartphones or tablets within two years.