Frontier Computing

Researchers and companies training biological neural tissue (wetware) today are limited to roughly 1 million neurones. Past that geometric scale, cells sit too far from an oxygen supply and die, a constraint the company calls the vascularisation problem. This caps how large a biological neural net can get, keeping wetware systems far smaller than what GPU-based training can reach and limiting what continual, in-substrate learning can be used for.

Frontier Computing has developed a tissue-culture approach intended to break the vascularisation limit, allowing biological neural nets to grow beyond the scale that was previously reasonable for this kind of tissue. The resulting systems carry continual learning built into the biological substrate itself, along with the sample efficiency associated with living neurones, and are paired with a training platform described as sitting an abstraction layer above backprop. The company provides the substrate, the interfacing, and the cloud infrastructure so that third parties can train on its systems. An early demonstration shows its neuronal tissue learning to play Frogger, reaching a 92% cross rate after one hour of realtime learning.