Section 7.2.5, “Learning in stages: the plateau and subsequent improvement”The plateau and rise suggest a temporary learning bottleneck; its cause is not established
Empirical result Supported by saved experimental runs, not by a proof.
1 Paper text
Learning in stages: the plateau and subsequent improvement. The plateau near and later rise in Figures 5–6 suggest a temporary learning bottleneck even with two-frame input. The learner makes substantial progress, remains at an intermediate level, and then improves sharply. This is evidence about training dynamics within the saved run, rather than a further circuit lower bound.
Generated from paper_jmlr.tex, lines 1126–1127.
2 Explain this
Even with the transition in the input, learning goes in stages: fast progress, a plateau near 0.54, then a sharp rise. The paper reads this as evidence about training dynamics in this one run, not as a further circuit lower bound.
3 How this is formalised
How this is covered. An empirical observation of the saved run. Its cause is not established, in the paper or here.
| Paper anchor | par-learning-in-stages-the-plateau-and-subsequent-improvement-p1 (website anchor; the manuscript has no label here) |
|---|---|
| Displayed as | Section 7.2.5, “Learning in stages: the plateau and subsequent improvement” |
| Claim | The plateau and rise suggest a temporary learning bottleneck; its cause is not established |
| Coverage category | Empirical result |
| Statement source | paper_jmlr.tex lines 1126–1127; SHA-256 of the normalised block source f1dd8351bd9a64ba… |
| Explanations | explain/c-plateau, formal/c-plateau |
| Build evidence | fresh isolated rebuild, audit report |
4 Lean statement and source
No Lean statement corresponds to this claim. This is an empirical statement about saved experimental runs. Lean proofs do not certify training runs, curves or checkpoints.
Build evidence: fresh isolated rebuild of all 38 bundled modules with Lean (version 4.26.0, started 2026-10-10T15:21:56Z, finished 2026-10-10T15:27:00Z: 38/38 modules compiled, 1588 axiom reports, no errors. See Build for the commands and the independent audit of 2026-10-10 (dated evidence).