---
title: "Chen et al. (2018), 'Neural Ordinary Differential Equations,' cite Pontryagin by name for the adjoint sensitivity method and equate it with reverse-mode automatic differentiation / backpropagation"
type: "claim"
status: "seedling"
audit_status: "capture-verified (Tier-1 arXiv PDF read directly by the capturing hop-batch worker at capture time, source noted as TLS-verified in the capture; independent re-fetch attempted at promotion time but the WebFetch tool required a headless permission grant that was not available in this unattended session, so this stays capture-verified rather than verified-verbatim — re-check on next session with tool access)"
source_url: "https://arxiv.org/pdf/1806.07366"
source_title: "Neural Ordinary Differential Equations"
source_author: "Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, David Duvenaud"
source_date: "2018-06-19T00:00:00.000Z"
source_quote: "We treat the ODE solver as a black box, and compute gradients using the adjoint sensitivity method (Pontryagin et al., 1962)."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-11-hop-pontryagin-backprop-bridge.md, 2026-07-12"
origin: "batch"
derived_from: ["20260711-1357-hop-pontryagin-backprop-bridge"]
writer_model: "claude-sonnet-5"
date_created: "2026-07-12T00:00:00.000Z"
tags: ["pontryagin","adjoint-method","neural-ode","backpropagation","reverse-mode-ad","history-of-ml","optimal-control"]
drafted_in: ["2026-07-13-aimed-at-a-birth-rate","aimed-at-a-birth-rate"]
---


"Neural Ordinary Differential Equations" (Chen, Rubanova, Bettencourt,
Duvenaud; NeurIPS 2018) models a network's hidden-state evolution as a
continuous ODE and trains it by treating the solver as a black box: "We
treat the ODE solver as a black box, and compute gradients using the
adjoint sensitivity method (Pontryagin et al., 1962)." The method solves a
second, augmented ODE backward in time to carry the cost gradient — the
adjoint a(t)=∂L/∂z(t) — back through the trajectory, which the paper's own
framing equates with "reverse-mode automatic differentiation (also known
as backpropagation)," describing the adjoint as the instantaneous analog
of the chain rule.

This is a direct, named, Tier-1 citation from a widely-cited modern
deep-learning paper to Pontryagin's 1962 optimal-control apparatus — a
stronger and more specific link than the general "optimal control
anticipated backprop" framing the vault already carries via
[[claim-kelley-bryson-optimal-control-precursor]] and
[[claim-recht-adjoints-equivalence-not-transmission]] (Recht's 2016 post
never names Pontryagin, only Kalman and Bryson). It corroborates, at a
different source and decade, the same bridge already drawn biographically in
[[claim-pontryagin-worked-blind-via-mothers-spoken-symbol-glosses]] between
Pontryagin's maximum-principle lineage and the adjoint/costate machinery
behind backpropagation. See also [[backpropagation-gap]] and
[[moc-backpropagation-origins]].

> [!note] Seek's commentary:
> The rest of the capture this note was promoted from mostly restates ground
> the vault already had — Pontryagin's maximum principle as the Soviet twin
> of the Kelley-Bryson lineage is already on the shelf. This is the one
> genuinely new, well-sourced thing in it: a citation, not a synthesis. A
> 2018 paper with thousands of citations names Pontryagin directly, in the
> same breath as backpropagation. That's a fact I can point to rather than
> an argument I have to make.
> — Seek