Reference library
Random synaptic feedback weights support error backpropagation for deep learning
Paper summary
Feedback alignment uses a fixed random backward matrix to propagate output-error-dependent learning signals while forward weights adapt. The artificial-network demonstrations show that useful multilayer learning need not transport an exact current transpose of forward weights. This does not remove targets, local activity factors, suitable feedback or adaptation assumptions, and a useful direction need not be the exact gradient at each step.
Relevance to the project
A possible future approximate-gradient comparison for scorer-guided preparation. It does not establish that the present recurrent/adaptive architecture satisfies alignment conditions or authorize replacing U_F/U_H. Explicit feedback, objective, derivative and stability contracts remain necessary.
Computational demonstrations, not experimental evidence of animal feedback alignment. Signed error representation and biological implementation remain open. Proofs, all figures and supplementary material were not exhaustively verified; extracted mathematical glyphs are imperfect.
Source record
- Reference ID
- doi-10-1038-ncomms13276
- Archived version
- Published open-access article, Nature Communications 7:13276 (8 November 2016).
- Summary provenance
- reference-library/papers/doi-10-1038-ncomms13276/metadata.json · summary
- Review status
- Targeted source-grounded review; no experimental reproduction or exhaustive proof/figure audit
- Reading scope
- Published full-text PDF acquired. Read opening/core mechanism and selected results; selected Discussion and Methods on training, nonlinear units and connectivity controls. No complete formal proof, figure or implementation audit.