Neural networks could run on native 'neuristors' in the future
A combination of capacitors with a new element called memristor, currently under development, will allow the construction of computational elements that work with neuron-like spikes, miniaturized on a chip. The authors of a study published in "Nature Materials" in December describing such a device named it "neuristor".
Neurons encode information in the pattern and timing of spikes (sharp signal pulses). The researchers used a simplified model of neurons based on sodium-potassium ion channels to turn the neuron on and off.
Each unit consists of a capacitor (to allow it to build up charge) in parallel to a memristor (which allows the charge to be released suddenly). The combination produces spikes of activity as soon as a given voltage threshold is exceeded.
Until now, neural networks run on neurons which are simulated by software. Building chips with massive amounts of neuristors on them could enable researchers to run and analyze neural networks and corresponding applications and experiments more efficiently.
HP labs is developing the underlying memristor technology, but apparently pushed back a planned date for production and release from 2013 to earliest end of 2014, for marketing reasons. Techspot cited Stan Williams from the Memristor Research Group at HP saying that the memristor technology would cut into the flash-memory business of production partner Hynix, since it allows superior speeds. Japanese researchers showed an example that could allow 11 times faster write-speeds in SSDs.