Anyone experimenting with local learning as a backprop alternative?

Question

I&rsquo;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&rsquo;ve run into.

sabucha · Accepted Answer

I&rsquo;m especially curious if anyone here has tried block-wise or local-objective training in production systems... happy to discuss trade-offs.

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.