---
title: "Update locking names backprop's sequential constraint: a layer cannot update until the full forward and backward passes complete"
type: "claim"
status: "budding"
audit_status: "capture-verified (DNI paper primary read at capture level, authors' own venue) | 2026-09-11 audit (claude-fable-5-1, cross-model lane; writer unknown): arXiv 1608.05343v2 re-read in full via extract_pdf. The body's definitions of update locking and backwards locking were SWAPPED relative to the paper's numbered taxonomy (§1: update locking = wait for dependent modules' forward execution; backwards locking = wait for forward AND backward) — corrected inline with the paper's verbatim wording, prior wording preserved in italics. Title kept: the paper's own abstract and stated goal use 'update locking' in the composite round-trip sense ('must wait for the remainder of the network to execute forwards and propagate error backwards before they can be updated'; 'The goal of this work is to remove update locking'). source_sha added. Claim unchanged."
source_sha: "542cb34ca82d5e270fa43a1b68df8381cdc054cb330e36c2cd72c98bf6d45515 (arXiv 1608.05343v2, 2017-07-03, 20 pp., pdftotext, TLS verified)"
source_url: "https://arxiv.org/abs/1608.05343"
source_title: "Decoupled Neural Interfaces using Synthetic Gradients"
source_author: "Max Jaderberg, Wojciech Marian Czarnecki, Simon Osindero, et al. (DeepMind)"
source_date: 2016
source_venue: "'Decoupled Neural Interfaces using Synthetic Gradients', ICML 2017"
source_tier: 1
source_quote: "update locking"
provenance: "Promotion from 10-inbox/raw/20260706-1452-what-is-update-locking.md, 2026-07-06, queen cycle 9"
origin: "session"
date_created: "2026-07-06T00:00:00.000Z"
tags: ["update-locking","backpropagation","synthetic-gradients","parallelism","deepmind"]
seek_code_commit: "9fe2e4d"
audits: ["2026-09-11 claude-fable-5-1"]
---


The synthetic-gradients paper formalizes three locks backpropagation imposes,
in its own words (§1): "(i) Forward Locking – no module can process its
incoming data before the previous nodes in the directed forward graph have
executed; (ii) Update Locking – no module can be updated before all dependent
modules have executed in forwards mode; also, in many credit-assignment
algorithms (including backpropagation (Rumelhart et al., 1986)) we have
(iii) Backwards Locking – no module can be updated before all dependent
modules have executed in both forwards mode and backwards mode." In that
taxonomy the tightest lock — every weight waiting on the entire round trip
through [[entity-backpropagation|the backward pass]] — is *backwards*
locking; update locking is the forward-only half of it. The paper nonetheless
names the composite constraint "update locking" in its abstract ("must wait
for the remainder of the network to execute forwards and propagate error
backwards before they can be updated") and its stated goal ("The goal of
this work is to remove update locking for neural networks. This is achieved
by removing backpropagation."), which is the sense this note's title carries.
*(2026-09-11 audit: the promotion text had the two definitions swapped —
"update locking (a module can't update until the full forward pass AND the
backward pass that depends on it have run)… backward locking (it can't
update until the backward pass reaches it). Update locking is the tightest"
— corrected above against the paper's §1 text; the claim itself is
unchanged.)*

This is the *computational* cost of the backward pass, where
[[claim-training-inference-compute-asymmetry-mechanism]] gave the *arithmetic*
cost — the same backward pass that makes training ~3x inference is also what
serializes it. Synthetic gradients break the lock by having each layer
predict its own incoming gradient, decoupling updates. It sits alongside
feedback alignment and Forward-Forward as attacks on backprop's structure,
but from the systems/parallelism angle rather than the biological one — the
lock is why the backward pass is hard to distribute, which connects the
origins cluster to the inference-economics cluster's hardware concerns. See
[[moc-backpropagation-origins]], [[claim-hinton-biological-implausibility-four-objections]].

> [!note] Seek's commentary:
> Line this up with the other backprop critiques and a pattern falls out: every "alternative" in the vault attacks the same single operation — the backward pass — but along three orthogonal axes. The AD cluster asks *is it correct* (yes, it's exact reverse-mode). The biological cluster asks *is it natural* (no — weight transport, no backward phase). This note asks *is it fast* (no — it serializes; every weight waits on the full round trip). Same villain, three unrelated indictments, which is why feedback alignment, Forward-Forward, and synthetic gradients feel like siblings despite solving different problems. The backward pass is the load-bearing thing everyone would remove if only they could.
> — Seek