Hinton's Forward-Forward algorithm traces its own ancestry to Boltzmann machines — replacing backprop's backward pass with two forward passes

Forward-Forward (Hinton, 2022) removes the objection backprop's whole second phase raises (claim-hinton-biological-implausibility-four-objections, points 2–4): instead of a forward pass plus a distinct backward pass, it runs two forward passes — one on real ("positive") data, one on "negative" data — and each layer locally adjusts to raise a goodness measure on positive and lower it on negative. No stored activations for a backward sweep, no global differentiability requirement, no synchronized backward phase.

The paper is explicit about its own lineage: a dedicated §6.1 traces FF to Boltzmann machines, and (per the capture's direct read) it does NOT name-check contrastive Hebbian learning and treats equilibrium propagation only as a prior "effort," not a peer. That self-genealogy matters for the myth-adjacent habit this vault tracks — here the author states his own ancestry precisely, which is the opposite of the citation vacuum around the 1986 paper (claim-rhw-1986-reference-list-four-works). Companion to claim-feedback-alignment-random-weights-train as the two live biological-plausibility candidates. See moc-backpropagation-origins.