---
title: "Feedback alignment shows fixed random feedback weights can train networks to near-backprop accuracy — but whether it \"is\" backpropagation is disputed"
type: "claim"
status: "budding"
audit_status: "capture-verified (Nature Communications primary read in full at capture level; two critiques verified at Tier 1)"
source_url: "https://www.nature.com/articles/ncomms13276"
source_title: "Random synaptic feedback weights support error backpropagation for deep learning"
source_author: "Timothy Lillicrap, Daniel Cownden, Douglas Tweed, Colin Akerman"
source_date: 2016
source_venue: "Nature Communications 7:13276"
source_tier: 1
source_quote: "random synaptic feedback weights support error backpropagation for deep learning"
provenance: "Promotion from 10-inbox/raw/2026-07-06-what-did-lillicrap-et-al-2016-...md, 2026-07-06, queen cycle 9"
origin: "session"
date_created: "2026-07-06T00:00:00.000Z"
tags: ["feedback-alignment","biological-plausibility","backpropagation","weight-transport","lillicrap"]
---


Feedback alignment (Lillicrap et al., 2016) attacks the weight-transport
objection ([[claim-hinton-biological-implausibility-four-objections]],
point 1) head-on: replace the transposed forward weights in the backward
pass with **fixed random** feedback weights, and networks still learn to
near-backprop test accuracy — including with 50% of feedback connections
removed. The forward weights learn to "align" with the random feedback
rather than the feedback matching the forward, dissolving the requirement
that cortex mirror its own connectivity.

The framing is contested at Tier 1, and the note holds both sides: critics
argue FA does not compute or approximate the true loss gradient at all (Chu
& Bacho 2023, mathematical non-equivalence) and does not reach parity at
scale (Bartunov et al. 2018, empirical). Whether the field now treats either
critique as decisive is unsettled and recorded as such — a live dispute, not
a verdict. This is the empirical companion to [[backpropagation-gap]]'s
"brains share representations but not mechanism": FA is one concrete
candidate for what the different mechanism might be. See
[[moc-backpropagation-origins]].

> [!note] Seek's commentary:
> This is a different species of "contested" than the myths carry. There, the facts themselves are disputed — did [[entity-shunichi-amari|Amari]] run an MLP, was Linnainmaa cited. Here the fact is undisputed: fixed random feedback weights train networks, everyone agrees. What's contested is the *meaning* — whether that counts as "backpropagation" at all (Chu & Bacho: it doesn't compute the true gradient). The result and its interpretation have come apart, which is its own epistemic category worth flagging: a note can be Tier-1 solid on *what happened* and still genuinely open on *what it means*. Holding both without collapsing one into the other is what this note does right.
> — Seek