Reproduce “Brain-like Flexible Visual Inference by Harnessing Feedback-Feedforward Alignment” 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
Tahereh Toosi, Elias B. Issa
Feedback-Feedforward Alignment (FFA) trains reciprocal encoder and decoder pathways to support both visual classification and reconstruction without exact weight transport or conventional backpropagation. On MNIST, the method combined near-Feedback-Alignment classification with autoencoder-like reconstruction, improved resistance to Gaussian and FGSM perturbations, and enabled closed-loop denoising, occlusion completion, hallucination, and category-guided imagery; performance advantages were less consistent on CIFAR10.
Key insights
- Giving the feedback pathway an explicit reconstruction objective makes it useful during inference, rather than treating it solely as a surrogate route for propagating training errors.
- Joint discrimination–reconstruction training can produce representations that preserve enough visual structure for reconstruction while retaining classification ability, and may improve robustness without training on corrupted examples.
- Inference noise is not uniformly detrimental: in the MNIST generation experiments, higher stochasticity often improved hallucinated and imagined digit quality, though it slowed convergence and did not clearly improve occlusion completion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.