San Francisco-based Rain Neuromorphics is developing an analogue chip for machine learning that is based on brain architecture, reports EE Times. Naturally, with a background in both ML and Neuroscience, this piqued my interest and if you are interested in ML hardware and/or the brain you should take a look, too. A bit off-topic here, but too cool for me to pass up the opportunity to tell you about this, just in case you find it as fascinating as I do.
Unlike conventional processors that rely on packing ever more transistors onto less and less silicon to increase processing power, an analogue brain-like chip is a radical design departure and holds out the promise of incredible performance and energy efficiency by dispensing with the intermediary of digital computation to represent network activity. Like the brain, the design has a base layer of neurons (c.f. the brain’s white matter) from which axons interconnected by dendrites extend (c.f. the brain’s grey matter). The interconnections are fixed according to a statistical pattern. The chips are trained by varying the strength of the interconnections, just like synaptic strength varying in the brain.
In addition to a brain-based design, Rain Neuromorphics is also working on training algorithms that work with this architecture, since back-propagation, the work-horse of machine learning, is known not to work in such architectures (and indeed the brain). One such approach is called predictive coding and you can read about that and other brain-friendly learning rules here. Even cooler, this study has found that something like predictive coding is evolutionarily adaptive for good metabolic management of (real) neurons in the brain and found that individual neurons can indeed predict their activity 10-20ms into the future and respond to changing stimuli in a way that reduces the difference between predicted and actual activity.