Reproduce “Feedforward Initialization for Fast Inference of Deep Generative Networks is biologically plausible” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yoshua Bengio, Benjamin Scellier, Olexa Bilaniuk, Joao Sacramento, Walter Senn
The authors show mutual prediction between bottom-up and top-down inputs can initialize a recurrent generative network near a fixed point through a feedforward pass. In an MNIST simulation, autoencoder-trained weights produce update steps that shrink sooner than random weights at both tested hidden-layer widths. This indicates faster convergence there, not a wall-clock speedup or a guarantee for other networks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.