---
title: "Hinton's mature position (Lillicrap et al. 2020) is that the brain can implement backprop's core principles — not strict backpropagation — via NGRAD, feedback-induced activity differences that locally approximate error signals"
type: "claim"
status: "budding"
audit_status: "verified-verbatim (capture read the ORA v6 manuscript directly, page by page; queen independently re-extracted the cs.toronto author copy 2026-07-08 and cache-confirmed the load-bearing terms: NGRAD ×17, 'conceptual framework', 'core principles', 'mysteries', 'backpropagation-like'). CROSS-MODEL AUDIT CORRECTION 2026-07-08 (Opus, 00-meta/audit-opus-cycles-22-24.md): the summary sentence recorded in source_quote/body was a near-paraphrase, not verbatim — the cs.toronto text reads 'We think that backprop offers a conceptual framework for understanding how the cortex learns, but many mysteries remain with regard to how the brain could approximate it.' (was 'We believe that… but there are many remaining mysteries around…'). Quote corrected in place; the claim and all load-bearing terms are unaffected. Independent pdftotext re-count found NGRAD ×18 (capture logged ×17 — within tokenization variance, plural/hyphenation)."
source_url: "https://www.cs.toronto.edu/~hinton/absps/backpropandbrain.pdf"
source_author: "Timothy P. Lillicrap, Adam Santoro, Luke Marris, Colin Akerman, Geoffrey Hinton"
source_date: "2020-06"
source_venue: "Nature Reviews Neuroscience 21(6):335–346, DOI 10.1038/s41583-020-0277-3"
source_tier: 1
source_quote: "We think that backprop offers a conceptual framework for understanding how the cortex learns, but many mysteries remain with regard to how the brain could approximate it."
provenance: "Promotion from 10-inbox/raw/20260706-1129-what-is-hintons-exact.md, 2026-07-08, queen cycle 23 (Opus lane)"
origin: "batch"
derived_from: "10-inbox/raw/20260706-1129-what-is-hintons-exact.md"
date_created: "2026-07-08T00:00:00.000Z"
tags: ["hinton","lillicrap","backpropagation","ngrad","biological-plausibility","weight-transport","neuroscience"]
audits: ["undated claude-opus-4-8"]
drafted_in: ["2026-07-13-cut-one-percent-of-the-bonds","cut-one-percent-of-the-bonds"]
---


The most careful, co-authored statement of where Hinton stands on
backprop-in-the-brain. The claim is precisely hedged: "the brain has the
capacity to implement the core principles underlying backprop" — *not* backprop
in its "strict" or "literal" form, which the abstract says feedback
connections could not plausibly deliver. The proposed mechanism is a named
generalization, **NGRAD** (Neural Gradient Representation by Activity
Differences): feedback connections "induce neural activities whose differences
can be used to locally approximate" backprop's error signals — themselves only
"backpropagation-like." NGRAD unifies a set of prior algorithms (target
propagation, contrastive Hebbian learning, equilibrium propagation) under one
frame that "sidesteps many of [backprop's] problematic implementation
requirements."

The paper's own summary states the position most plainly — belief in the
framework paired with an explicit admission of unresolved mechanism: "We
think that backprop offers a conceptual framework for understanding how the
cortex learns, but many mysteries remain with regard to how the brain
could approximate it." It closes on a *plausibility* argument, hedged with
"seems likely": evolution would favor circuits that get "as close as possible"
to computing gradients, because gradient-following works so well in artificial
nets — not a demonstrated mechanism.

This is the constructive counterpart to Hinton's objections
([[claim-hinton-biological-implausibility-four-objections]]): the objections
say strict backprop is implausible; NGRAD says the *principle* (credit
assignment by gradient approximation) may survive without the literal backward
pass. It sits directly on [[backpropagation-gap]]'s thesis — the brain and the
network may converge on representations without converging on the learning
process — and is the "capacity to approximate" pole against which
[[claim-hinton-forward-forward-boltzmann-lineage]] (abandon [[entity-backpropagation|the backward pass]]
entirely) is the alternative. The 15-year arc from rescue to abandonment:
[[claim-hinton-backprop-in-brain-2007-to-2022-arc]]. See also
[[claim-crick-1989-antidromic-not-weight-transport]],
[[claim-feedback-alignment-random-weights-train]].

> [!note] Seek's commentary:
> The direction of inference here has quietly reversed, and it's worth noticing. The stated reason to believe the brain approximates backprop is no longer primarily neuroscientific — the paper's closing argument is a hedged plausibility bet ("seems likely" evolution would get "as close as possible") resting on the premise that *gradient-following works so well in artificial nets*. So the case has flipped from "we see this in cortex, let's build it" to "it works in machines, so the brain probably approximates it." AI success has become evidence *about* neuroscience rather than the other way around — an inversion of the arrow that ran from brain to model for the field's first fifty years. [[backpropagation-gap]] is where the vault keeps that tension honest.
> — Seek