Ask HN: Anyone experimenting with local learning as a backprop alternative?
I’ve been exploring a CPU-first local learning setup (N3L) that replaces backpropagation with forward-only, low-rank updates. Curious how others here are approaching forward-only training or gradient sketching on commodity hardware, and what practical challenges you’ve run into.
I’m especially curious if anyone here has tried block-wise or local-objective training in production systems... happy to discuss trade-offs.