Intel Loihi mimics human brain function for self-learning

Greg S

Posts: 1,607   +442

What seems like it should be straight out of a science fiction movie is actually found inside Intel's laboratories. In a partnership with the California Institute of Technology, Intel has developed the first ever self-learning neuromorphic chip.

Codenamed Loihi, the chip functions similar to a human brain by taking environmental feedback into consideration when executing instructions. Unlike traditional deep learning systems, Loihi is able to adapt and response in scenarios that are being seen for the first time. Intel indicates that such a system can be highly personalized for any end-user since only a small amount of training is required to achieve desired responses.

Cybersecurity applications are one potential use since any deviation from what Loihi determines to be a normal network state could cause an alert to be issued. Biomedical applications such as detecting irregular heartbeats specific to an individual were also mentioned as possible uses.

Loihi is built using a 14nm process and contains a total of 130,000 neurons with 130 million synapses. The new chip is claimed to be over 1,000 times more energy efficient compared to current general purpose compute hardware for system training. Artificial intelligence systems typically require large amounts of processing power but Intel has implied that micro-instances of AI for low-power applications could be implemented in IoT and mobile devices.

University researchers working with Intel will receive Loihi chips in the first half of 2018 to explore potential uses and improve AI technologies. A commercial release isn't planned anytime soon.

Permalink to story.



Posts: 506   +162
Just don't teach it to think like we do considering the predicament Humanity currently finds itself doing so may be somewhat akin to gimping the thing out. ;)


Posts: 63   +11
I wonder if this chip can exhibit a form of "curiosity" and want to know things. If that were true, then giving it an in system link to normal computing power so it could look things up would perhaps help with the learning process in a general way. In fact, combining the three types of computing now becoming available, quantum, neural, and good old digital in one package (or cyborg body?) might result in something like human intelligence. Sure, sure, I know that quantum at this time requires near absolute zero temperatures to operate cleanly, but hey... it could happen! Error checking on quantum operations using perhaps the other two forms of computing I mentioned might eliminate some of the noise and produce results even at "room temperatures" if the systems were robust enough. Humans filter out noise all the time. It's a learning process that might be machine producible.