2 ms·
> PC-ALM closes the PC-BP gap at matched inference budget in nonlinear networks Okay, so no efficiency improvement?
by erichocean 11d ago
> PC-ALM closes the PC-BP gap at matched inference budget
in nonlinear networks
Okay, so no efficiency improvement?
- Tuna-Fish 11d agoBackprop has significant costs, if a local solution truly matches its performance, it will be dropped like a hot potato.
- erichocean 11d agoAgreed, but how will we find out? Pre-training alone on frontier-class models costs billions of dollars.