Japanese Team Builds AI Computing System With Rat Neurons That Generates Complex Sequential Signals

A team from Tohoku University and Future University in Japan recently trained rat cortex neurons and built a 'closed-loop reservoir computing' system with a real-time machine learning framework. Without external input, the system can independently learn and generate periodic and chaotic waveforms to perform AI computing tasks.

The team used PDMS microfluidic films to constrain neuron connections, confining cell bodies in 128 microholes connected by microchannels to form lattice and hierarchical network structures. This significantly reduced inter-neuron correlation and improved network dynamic dimensions. Tests showed the lattice network performed well, generating various waveforms and approximating the Lorenz attractor, with a correlation of over 0.8 between predicted and target signals.

However, the technology has performance bottlenecks: errors increase when running independently after training stops, and feedback delay limits its ability to track fast-changing waveforms. The team plans to reduce delay via dedicated hardware and expand its application in brain-machine interfaces and neural prosthetic devices.