What did Lillicrap et al. (2016, Nature Communications) actually demonstrate with feedback alignment, and on what grounds do critics dispute that it constitutes backpropagation?

Short answer: Lillicrap, Cownden, Tweed & Akerman (2016) demonstrated that a deep network can be trained to backprop-competitive accuracy using fixed, random backward weights instead of the transpose of the forward weight matrix — removing the "weight transport" requirement that makes standard backpropagation biologically implausible. This is well-supported directly from the primary paper (Tier 1). Whether this "constitutes backpropagation" is disputed on two distinct, independently sourced grounds: (1) a mathematical objection — a 2023 analysis argues the feedback-alignment update rule is not a gradient of any loss function and does not approximate backpropagation's gradient descent at all, characterizing it instead as a random walk that happens to be attracted toward the same fixed points BP would find; and (2) an empirical scaling objection — a 2018 NeurIPS paper, co-authored by Lillicrap himself, found that feedback alignment (and the related target-propagation family) perform significantly worse than backpropagation on harder benchmarks (CIFAR-10, ImageNet) and in more structured (locally-connected/convolutional) architectures, unlike the near-parity Lillicrap et al. reported on their original, simpler tasks.

Claim 1: The paper's core demonstration — random, fixed feedback weights train networks to backprop-competitive accuracy

Claim type: Specific technical-mechanism claim → Tier 1–2 required. Source tier: Tier 1 (Nature Communications 7:13276, primary paper, full text).

The paper's abstract states its central finding directly:

"In machine learning, the backpropagation algorithm assigns blame by multiplying error signals with all the synaptic weights on each neuron's axon and further downstream. However, this involves a precise, symmetric backward connectivity pattern, which is thought to be impossible in the brain. Here we demonstrate that this strong architectural constraint is not required for effective error propagation. We present a surprisingly simple mechanism that assigns blame by multiplying errors by even random synaptic weights. This mechanism can transmit teaching signals across multiple layers of neurons and performs as effectively as backpropagation on a variety of tasks."

Mechanistically, standard backpropagation multiplies the error signal e by the transpose of the forward weight matrix, Wᵀ, to produce the backward-propagated error at each layer. Feedback alignment (FA) instead uses a fixed, randomly-initialized matrix B in that role: δ_FA = Be. B is never updated to track W; it is chosen once, at random, and left fixed throughout training. The paper terms this the resolution of the "weight transport problem" — the requirement, otherwise unmet by any known cortical anatomy, that feedback connections carry information identical to the feedforward weights.

Source: Lillicrap, T.P., Cownden, D., Tweed, D.B., Akerman, C.J. (2016). "Random synaptic feedback weights support error backpropagation for deep learning." Nature Communications 7, 13276. Published 2016-11-08. Full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC5105169/. Publisher page: https://www.nature.com/articles/ncomms13276.

Wikilink note: See also backpropagation-gap, weight-transport-problem, claim-linnainmaa-reverse-mode-single-pass.

Claim 2: Quantitative results — FA matched backprop's test error on the tasks tested, including with 50% of feedback connections removed

Claim type: Quantitative claim → Tier 1–2 required. Source tier: Tier 1 (same primary paper, full text).

On a hand-written digit classification task, the paper reports:

"On this task, backprop brings the mean error on the test set to 2.4%, averaged over 20 runs. Feedback alignment learns just as quickly, achieving 2.1% mean error."

On a robustness test that removed most of the reciprocal wiring between the random feedback matrix and the forward weights:

"Even when we randomly remove 50% of the elements of the W and B matrices, so that neurons in h and y have a 25% chance of reciprocal connection, feedback alignment still matches backprop (2.4% mean error; n=20)."

And on a four-layer function-fitting task:

"Tested on a function fitting task with a four-layer network, feedback alignment performed as well as backprop."

A separate deeper network reportedly reached "1.8% final error on the test set" using feedback alignment alone.

Source: same as Claim 1, PMC full text at https://pmc.ncbi.nlm.nih.gov/articles/PMC5105169/, accessed 2026-07-06.

Claim 3: Critics dispute the "constitutes backpropagation" framing on mathematical grounds — FA is argued to not compute (or approximate) the loss gradient at all

Claim type: Specific technical-mechanism claim → Tier 1–2 required. Source tier: Tier 1 (arXiv preprint, named academic authors, full mathematical derivation read directly).

Chu & Bacho (2023), in a paper investigating why FA's update rule "aligns" with the true gradient at all, describe the mechanism precisely and then dispute the standard interpretation. On the mechanism:

"FA is the same as BP, except that the terms ∂h(·)_i/∂f(·)_j appearing in [the backprop chain-rule term] are replaced by randomly chosen (but fixed) numbers Rij... Note, that the rhs of the equation is not a gradient of a particular function."

On the central dispute — whether the "gradient alignment" that FA exhibits means it is approximating backpropagation:

"In the literature this alignment is commonly assumed to be the driver for the performance of the FA algorithm. It remains unclear, however, why FA aligns... Our theoretical results suggest a somewhat different view: FA is not approximating BP at all, and indeed does not descend or ascend the gradient of a loss function or an approximation thereof. It is not 'learning' in the sense one normally understands this term. Instead, it performs a random walk in weight space."

Their proposed alternative account: FA's output layer genuinely does gradient descent (in exactly the same way as BP, since the last layer's update happens to be independent of the random feedback matrix), which pulls the whole network toward points in weight-space that are also fixed points of the FA update rule; the paper's contribution is showing that "gradient alignment" is merely the stability condition for a network staying near one of those fixed points once it arrives — not the mechanism that drives the network there in the first place. They also connect this to a documented empirical weak point: "The often cited failure of FA for convolutional neural networks is likely a consequence of the sparseness of loss extrema for those networks" — i.e., the random-walk account explains, on their view, why FA is known to work poorly for CNNs specifically.

Source: Chu, D. & Bacho, F. (2023). "Random Feedback Alignment Algorithms to train Neural Networks: Why do they Align?" arXiv:2306.02325 [cs.LG], submitted 2023-06-04. Full text read via extract_pdf, https://arxiv.org/pdf/2306.02325.

Claim 4: Critics dispute the "constitutes backpropagation" framing on empirical-scaling grounds — FA underperforms BP significantly on harder tasks/architectures, including in a paper co-authored by Lillicrap himself

Claim type: Quantitative + technical-mechanism claim → Tier 1–2 required. Source tier: Tier 1 (NeurIPS 2018 peer-reviewed proceedings paper, abstract verified directly).

Bartunov, Santoro, Richards, Marris, Hinton & Lillicrap (2018) set out specifically to stress-test feedback alignment (and target propagation) on harder benchmarks and more structured architectures than the original 2016 paper used. The abstract states their motivating gap and central finding:

"As of yet, none of these proposals have been rigorously evaluated on tasks where BP-guided deep learning has proved critical, or in architectures more structured than simple fully-connected networks. Here we present results on scaling up biologically motivated models of deep learning on datasets which need deep networks with appropriate architectures to achieve good performance. We present results on the MNIST, CIFAR-10, and ImageNet datasets and explore variants of target-propagation (TP) and feedback alignment (FA) algorithms... Many of these algorithms perform well for MNIST, but for CIFAR and ImageNet we find that TP and FA variants perform significantly worse than BP, especially for networks composed of locally connected units, opening questions about whether new architectures and algorithms are required to scale these approaches."

That Timothy Lillicrap is a co-author of this 2018 paper is itself worth flagging: this is not an outside adversarial critique but a self-directed empirical audit, published two years after the original Nature Communications result, that narrows the original claim considerably — the 2016 near-parity result held for MNIST-scale, fully-connected settings, but did not generalize to the harder, more structured settings that matter for most practical deep learning.

Source: Bartunov, S., Santoro, A., Richards, B., Marris, L., Hinton, G.E., Lillicrap, T. (2018). "Assessing the Scalability of Biologically-Motivated Deep Learning Algorithms and Architectures." Advances in Neural Information Processing Systems 31 (NeurIPS 2018). Abstract page confirmed live: https://papers.nips.cc/paper_files/paper/2018/hash/63c3ddcc7b23daa1e42dc41f9a44a873-Abstract.html.

Further leads

• Lillicrap, Santoro, Marris, Akerman & Hinton (2020), "Backpropagation and the brain," Nature Reviews Neuroscience 21(6), 335–346 — a later co-authored review (with Hinton) revisiting FA's status; already flagged in the 2026-06-30 capture on Hinton's biological-implausibility argument as paywalled/unquoted in this vault so far. Would be the natural next stop to see how Lillicrap himself, in his own later words, frames the "does this constitute backprop" question with the benefit of both the 2018 scaling result and independent critiques like Chu & Bacho in hand.
• Refinetti, D'Ascoli, Ohana & Goldt (2021), "Align, then memorise: the dynamics of learning with feedback alignment" (ICML) — cited within Chu & Bacho as the source of the "vanishing initial weights → stronger alignment" mechanism; not independently read this session, would be worth fetching directly for its own exact claims rather than relying on Chu & Bacho's characterization of it.
• Nøkland (2016), direct feedback alignment (DFA) — the layer-parallel variant of FA; Chu & Bacho's stability-criterion argument is explicitly extended to DFA in their paper, but the original Nøkland paper itself was not read this session.
• Whether Grossberg (1987) or Crick (1989) originated the "weight transport problem" terminology used throughout this literature was not independently re-verified this session (surfaced only in unverified WebSearch synthesis, not a directly quoted primary source); the 2026-06-30 Hinton capture already flags Crick (1989) as an open lead for the same reason. [unsourced — needs verification, dropped from claims above].

Central question status: The core empirical question ("what did Lillicrap et al. 2016 demonstrate") is confirmed directly from the primary paper at Tier 1 — fixed random feedback weights train networks to near-parity with backprop on the tasks tested, including under 50% connection removal. The dispute over whether this "constitutes backpropagation" is also confirmed at Tier 1 on two independent grounds (mathematical non-equivalence per Chu & Bacho 2023; empirical non-parity at scale per Bartunov et al. 2018). Not resolved, and not expected to be: whether the field currently regards either critique as decisive, or whether a consensus reading has emerged since 2023, was outside what this session's searches could establish — the two critiques are recorded as live, sourced disputes rather than as a settled verdict.