The Most Counterintuitive Way to Build a Brain
In this video, we explore Reservoir Computing: a radical approach to recurrent neural networks inspired by how biological brains might actually work. Instead of precisely engineering every connection in the network to produce perfectly-tuned dynamics, we leave a random bucket of neurons untouched and simply learn to “listen” to it in the right way, reducing a complex learning problem to linear regression. The secret behind why this works turns out to be deeply connected to Fourier’s 200-year-old insight about building any signal from simple building blocks.
FWIW, he makes a shoutout to TBT about six minutes in.