Comment by mapontosevenths
9 hours ago
What a cool idea. How does it work? AFAIK The human brain at least does sparse backprop and has SOME neural circuits that feed-backward, so how do you manage it without anything?
I tinkered for a minute but never got anywhere.
Thanks! I have other ideas, following Jeff Hawkins's Thousand Brains Project, but in this one I'm trying to get to cortical columns from the other side, from "standard" deep neural networks.
The short version: each layer trains itself independently using Hinton's Forward-Forward algorithm. Instead of propagating error gradients backward through the whole network, each layer has its own local objective: "real data should produce high activation norms, corrupted data should produce low ones." Gradients never cross layer boundaries. The human brain is massively parallel and part of that is not using backprop, so I'm trying to use that as inspiration.
You're right that the brain has backward-projecting circuits. But those are mostly thought to carry contextual/modulatory signals, not error gradients in the backprop sense. I'm handling cross-layer communication through attention residuals (each layer dynamically selects which prior layers to attend to) and Hopfield memory banks (per-layer associative memory written via Hebbian outer products, no gradients needed).
The part I'm most excited about is "sleep". During chat, user feedback drives reward-modulated Hebbian writes to the memory banks (instant, no gradients, like hippocampal episodic memory). Then a /sleep command consolidates those into weights by generating "dreams" from the bank-colored model and training on them with FF + distillation. No stored text needed, only the Hopfield state. The model literally dreams its memories into its weights.
Still early, training a 100M param model on TinyStories right now, loss is coming down but I don't have eval numbers yet.
Neat. That thousand brains site looks right up my alley. If you haven't seen it, maybe check this out: https://www.nature.com/articles/ncomms13276
The idea is that the brain uses what the authors refer to as "feedback alignment" rather than backprop. Even if it turns out not to be literally true of the brain, the idea is interesting for AI.
I also love the idea of grafting on the memory banks. It reminds me of early work on DNC's (Differentiable Neural Computer's). I tried to franken-bolt a DNC onto an LLM a few years back and mostly just earned myself headaches. :)
It's fun to see all the wild and wacky stuff other folks like myself are tinkering with in the lab.