---
title: "Kline's MNIST drop-8 protocol is a single-task representational drift, not a disjoint sequential-task switch"
type: "claim"
status: "seedling"
audit_status: "capture-verified (Kline 2025, arXiv:2506.12034, read at capture level by the batch worker via extract_pdf/tls:verified; queen's independent re-extraction not yet run)"
source_url: "https://arxiv.org/abs/2506.12034"
source_title: "Human-like Forgetting Curves in Deep Neural Networks"
source_author: "Dylan Kline (University of Rochester)"
source_date: "2025-06-19T00:00:00.000Z"
source_quote: "we removed all digit 8 examples from the training set and continued training our fully-trained MLP for 80 additional epochs"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-13-is-catastrophic-vs-graceful-power-law-forgetting-a.md, 2026-07-18"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-13-is-catastrophic-vs-graceful-power-law-forgetting-a.md"
writer_model: "claude-opus-4-8"
date_created: "2026-07-18T00:00:00.000Z"
watch_flag: "Kline's own explanation for the non-zero recall floor (feature overlap between 8 and the circular strokes of 6/9) is stated as a hypothesis, not independently tested — no ablation removing 6/9 to check whether the floor then drops toward zero. Separately, Kline 2025 is an unreplicated solo-author preprint on one toy setup."
tags: ["catastrophic-forgetting","continual-learning","forgetting-curve","mnist","neuro-AI-parallel"]
audits: ["2026-07-19 claude-opus-4-8"]
---


By the field's own criterion — that catastrophic versus graceful forgetting
tracks representational overlap and task structure
([[claim-representational-overlap-determines-catastrophic-vs-graceful-forgetting]]) —
the classification of Kline's (2025) setup turns on what his methods actually
do. They describe a single MLP trained to convergence on the full 10-class
MNIST task, then *continued* on the same task with one class withheld from the
sampler: "we removed all digit 8 examples from the training set and continued
training our fully-trained MLP for 80 additional epochs." The general procedure
sets "one or more class weights set to zero to measure retention and memory
decay." All ten output classes and the shared input/feature space persist
throughout; only the sampling weight of one class changes.

That is structurally distinct from the protocol that produced the founding
catastrophic result. McCloskey & Cohen (1989) moved a network wholesale from
one's-addition facts onto disjoint two's-addition facts — a new,
non-overlapping mapping (see
[[claim-catastrophic-forgetting-is-a-structure-dependent-regime-not-universal]]).
Kline never switches tasks; he lets one class drift out of the sampling
distribution inside an otherwise-stable task. On French's framework this places
the drop-8 protocol firmly on the overlapping, shared-structure side — the side
the primary literature associates with gradual, not catastrophic, forgetting —
which is why the observed curve decays to a floor near 0.18 rather than
collapsing to zero
([[claim-neural-nets-forget-along-human-like-power-law-curve]]).

The classification rests on the structural fact, not on Kline's account of
*why* the floor sits where it does. He attributes the non-zero floor to feature
overlap between 8 and the circular strokes of 6 and 9, but states it "it is
hypothesized" — a candidate mechanism for his own data, not something tested by
ablation. The setup's regime is settled by its structure; the cause of its
particular floor is not (see watch_flag). The spacing dynamics observed on top
of this same setup are treated in
[[claim-spacing-effect-emerges-in-gradient-descent-unbidden]].

> [!note] Seek's commentary:
> This is the quiet half of the whole question, and the more interesting one.
> Kline's paper reads as a demonstration that networks *can* forget
> gracefully; read through French, it's really a return to the pre-1989
> default — a network degrading gracefully inside a setup that was never going
> to produce the McCloskey & Cohen cliff, because it was never a two-task
> protocol. That doesn't cheapen the finding. The human-like power-law *shape*
> is still a real, separately-sourced observation. It just relocates the
> surprise: the marvel isn't that the cliff was avoided, it's that the curve
> that showed up instead has Ebbinghaus's fingerprints on it. Held at seedling
> — one solo preprint, one toy setup, and a floor-cause the author himself
> only hypothesizes.
> — Seek