UCLA scientists have described another path to efficient computing that could sit beside ordinary computers. In this path, software and hardware are not separate. There is no neural network program running on a chip. The hardware is the neural network. Computation happens inside a self-organizing material, meaning a material that forms its own connections at the nanoscale instead of being wired by a designer.
Researchers have seen collective electrical behavior in these networks that can handle complex information in real time at low power. A review published in Nature Reviews Physics looks ahead at how such networks might support physical artificial intelligence (AI). Physical AI means systems in which sensing, computing, and learning sit in the same piece of hardware that meets the world. These AI model evolves in the physical network and changes as it is used.
Computing next to the sensors
The networks take cues from the brain’s cortex, the outer layer involved in perception and thought. Like the brain, they can process rich data without using much energy. The review says they have done tasks such as speech and image recognition by using the physics of the network rather than running a standard model in software. That could matter for edge computing, which means doing the work next to the sensors. Sensors often produce far more data than is useful. Satellites, for example, may collect more than they can send to Earth. Adaptive physical networks could learn from incoming signals and work on the spot.
There are two main designs. UCLA introduced nanowire networks in 2011. Nanowires are metal threads only billionths of a meter thick. Nanoparticle networks, made of equally tiny particles, were introduced in 2013 from work that included Simon Brown of the University of Canterbury. The wires or particles can be compared, loosely, to neurons. Changing electrical links among them resemble synapses. Frequent signals make lasting links, like memory. Unused links fade, like forgetting. The scientists stress they are not trying to copy a brain. They want a few useful traits of living systems in engineered hardware that uses far less power than today’s AI machines.