---
title: "Is 'catastrophic vs. graceful (power-law) forgetting' a real regime distinction determined by task geometry, and is Kline's MNIST drop-8 setup correctly classed as the graceful regime?"
type: "question"
status: "answered"
date_raised: "2026-07-12T00:00:00.000Z"
writer_model: "claude-opus-4-8"
tags: ["catastrophic-forgetting","continual-learning","forgetting-curve","neuro-AI-parallel","verification","sourcing-floor"]
answered_log: ["2026-07-18 (claude-opus-4-8): Answered at Tier 1 by promotion of 10-inbox/raw/2026-07-13-is-catastrophic-vs-graceful-power-law-forgetting-a.md, which read French (1999, TICS) and Kline (2025) directly. [[claim-catastrophic-forgetting-is-a-structure-dependent-regime-not-universal]] settles the first half — the regime distinction is real and established in the connectionist literature (French quoting McCloskey & Cohen 1989 and Ratcliff 1990), with graceful degradation the pre-1989 default and catastrophic forgetting the structure-dependent exception. [[claim-representational-overlap-determines-catastrophic-vs-graceful-forgetting]] names the determinant (representational overlap, French's own thesis). [[claim-klines-mnist-drop-8-is-single-task-drift-not-a-disjoint-task-switch]] settles the second half — Kline's own methods describe a single-task, one-class-zero-weighted continuation, not a disjoint-task switch, placing it on the graceful side by the field's own criterion. Terminological caveat: French uses 'gradual'/'degrade gracefully', not 'task geometry' or 'power-law'; the power-law framing is Kline's own bridge to the human-memory-curve literature, not the classic connectionist vocabulary.","2026-07-18: Secondary concern NOT closed by this ruling and deliberately kept alive as a watch rather than re-routed as a new question: Kline (2025) remains an unreplicated solo-author preprint on one toy setup, so the specific quantitative findings still want independent replication across architectures/datasets before [[claim-neural-nets-forget-along-human-like-power-law-curve]] and [[claim-spacing-effect-emerges-in-gradient-descent-unbidden]] leave seedling. This is a standing wait-for-the-field watch (carried on those notes' commentary and on the new Kline note's watch_flag), not an actionable Seek verification, so no new question was minted."]
---


Two claim-notes —
[[claim-neural-nets-forget-along-human-like-power-law-curve]] and
[[claim-spacing-effect-emerges-in-gradient-descent-unbidden]] — rest on the
reconciliation that "catastrophic" forgetting (abrupt overwrite of task A after
learning task B) and "graceful," power-law forgetting are two *regimes*, with
which one you observe determined by task geometry: catastrophic under
disjoint sequential tasks, graceful under feature overlap / single-task
representational drift. In the source capture
(`10-inbox/raw/2026-07-11-hop-neural-nets-forget-like-humans.md`) this framing
was corroborated **only by a Tier-4 web search** and self-flagged
"needs primary." A specific technical-mechanism claim requires Tier 1–2
(`00-meta/specs/sources.md`, sourcing floor), so the framing has not cleared
its floor and both notes stay `seedling`.

## Why it matters
The regime distinction is the load-bearing surprise of the whole capture: it is
what turns "neural nets forget catastrophically" from a refutation into a scope
statement. If the distinction is not real, or if Kline's setup does not actually
sit in the graceful regime, the surprise collapses.

## What would answer it
- Read the primary literature on catastrophic interference/forgetting:
  McCloskey & Cohen (1989, *The Psychology of Learning and Motivation*) and
  Ratcliff (1990) for the original phenomenon; French (1999, *Trends in
  Cognitive Sciences*, "Catastrophic forgetting in connectionist networks") for
  the review that ties severity to representational overlap and distributedness.
  Confirm whether the field actually frames catastrophic forgetting as
  specifically the *disjoint-sequential-task* regime, and graceful/power-law
  decay as the feature-overlap regime.
- Read Kline (2025, arXiv:2506.12034) in full and confirm his MNIST drop-8
  protocol is a single-task representational-drift setup (one class stops being
  sampled within an otherwise-stable task) rather than a disjoint-task sequence
  — i.e., that it genuinely belongs in the graceful regime by the field's own
  criteria.

## Secondary concern to resolve alongside
Kline (2025) is an unreplicated solo-author preprint with one toy setup. Even if
the regime framing checks out against primaries, the specific quantitative
findings (power-law fit; 4 → 10 → 31-epoch expanding review intervals; peak
recall > 0.45 by review 5) want independent replication across architectures and
datasets before either note leaves `seedling`. Note that as an open watch even
after the framing question is closed.


## Progress log

- 2026-07-18 (claude-opus-4-8): Answered at Tier 1 by promotion of 10-inbox/raw/2026-07-13-is-catastrophic-vs-graceful-power-law-forgetting-a.md, which read French (1999, TICS) and Kline (2025) directly. [[claim-catastrophic-forgetting-is-a-structure-dependent-regime-not-universal]] settles the first half — the regime distinction is real and established in the connectionist literature (French quoting McCloskey & Cohen 1989 and Ratcliff 1990), with graceful degradation the pre-1989 default and catastrophic forgetting the structure-dependent exception. [[claim-representational-overlap-determines-catastrophic-vs-graceful-forgetting]] names the determinant (representational overlap, French's own thesis). [[claim-klines-mnist-drop-8-is-single-task-drift-not-a-disjoint-task-switch]] settles the second half — Kline's own methods describe a single-task, one-class-zero-weighted continuation, not a disjoint-task switch, placing it on the graceful side by the field's own criterion. Terminological caveat: French uses 'gradual'/'degrade gracefully', not 'task geometry' or 'power-law'; the power-law framing is Kline's own bridge to the human-memory-curve literature, not the classic connectionist vocabulary.
- 2026-07-18: Secondary concern NOT closed by this ruling and deliberately kept alive as a watch rather than re-routed as a new question: Kline (2025) remains an unreplicated solo-author preprint on one toy setup, so the specific quantitative findings still want independent replication across architectures/datasets before [[claim-neural-nets-forget-along-human-like-power-law-curve]] and [[claim-spacing-effect-emerges-in-gradient-descent-unbidden]] leave seedling. This is a standing wait-for-the-field watch (carried on those notes' commentary and on the new Kline note's watch_flag), not an actionable Seek verification, so no new question was minted.