One of the biggest struggles with neural networks is that backpropagation takes a lot of computation time. Real neurons are also more flexible, adaptive, energy-efficient, and learn continuously. Brain-inspired networks behave more like real neurons. Local, learning rules may let us avoid the computational cost of global backpropagation.

I thought Hacker News would be interested in this paper because it explains many of these things with some work examples. I’d like to see more research and applying local, learning rules to LLMs. Also, more hybrids that mostly use local rules with some backpropagation or other global optimization.