acceptodds
Under review as a conference paper at ICLR 2027

PRINCIPAL CURVE NOVELTY GEOMETRY: FALSIFIABLE ATTRIBUTION OF NOVELTY IN FROZEN SEQUENTIAL REPRESENTATIONS

Abstract

Post-hoc novelty methods usually reduce representation unusualness to a single scalar, obscuring whether an observation is unusual because of its position along a learned trajectory, its deviation from that trajectory, or its rate of progression. We introduce principal-curve novelty geometry (PCNG), a post-hoc measurement con- tract for frozen sequential representations that asks which geometric interpretation is supported and when attribution should be refused. PCNG fits a training-only principal curve, tests whether the resulting progression geometry is admissible, and, when permitted, derives progression, transverse deviation and traversal rate on a common rarity scale calibrated against the training reference. Targeted interventions test these interpretations, while unit-preserving randomisation evalu- ates residual coupling beyond the dependence already present within trajectories. Across 117 runs in degradation prognostics spanning 13 frozen configurations, failure-mode interventions support the transverse interpretation in 35 of 36 di- rected tests, whereas the rate interpretation is unsupported. Cross-unit scrambling reduces the transverse hit rate from 0.972 to 0.667, while a within-unit control remains at 1.000. Residual coupling is heterogeneous, precluding an independence claim; the admissibility gate also varies across encoders and rejects some external- domain runs. Operationally, transverse rarity yields a 4.9% false-alarm rate versus 29.2% for a learned early-degradation health monitor at matched coverage, while distance-based novelty remains competitive. PCNG therefore contributes falsifiable attribution rather than universally better detection.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation Learning, Novelty Detection, Novelty Measurement, Predictive Novelty, Principal Curves, Progression Geometry, Frozen Representations, Novelty Attribution, Construct Validation, Prognostics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_omQTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes PCNG to analyze novelty in frozen sequential representations using a training-only principal curve. It derives progression position, transverse deviation, and traversal rate, calibrates them to rarity scores, and tests their interpretations. Experiments mainly support the transverse interpretation, while the rate interpretation remains unsupported. An admissibility gate determines whether attribution should proceed.

Critical strengths
  1. Testing the meaning of geometric scores is a useful goal. The paper reports negative results and distinguishes a computable score from an interpretation that is supported.
  2. The unit-preserving randomization addresses a statistical issue: correlated windows should not be treated as independent observations.
Critical weaknesses
  1. The evidence for novelty attribution beyond distance-based novelty remains limited. Transverse rarity is calibrated from distance to the curve, and the experiment yielding 35/36 changes only the reference distribution, with the curve and probes fixed. The controls and residual-scramble checks help, but it is still unclear how much the result validates the transverse interpretation beyond sensitivity to rarity calibration. The matched-Mahalanobis check in Appendix V is also PARTIAL.
  2. The admissibility gate needs stronger support as a basis for permitting or refusing attribution. CV(L) measures path homogeneity, but Table 20 shows construct accuracy declining before GO is withdrawn under branching, with limited sensitivity to curvature and transverse noise. Bootstrap gate-label agreement of 0.149 raises further concerns about stability. Loss of GO also does not imply refusal, since MARGINAL still permits attribution.
Questions for authors
  1. What do the reference-ablation interventions establish about the transverse interpretation beyond changes in rarity calibration, after accounting for the existing controls?
  2. Given the PARTIAL result in Appendix V, which findings support novelty attribution beyond distance-based novelty?
  3. How were the CV thresholds of 0.15 and 0.35 chosen? Does construct accuracy differ between GO/MARGINAL and NO-GO runs?
  4. With bootstrap gate-label agreement of 0.149, how should users decide whether to proceed, especially near the thresholds?
Reviewer_uMxeOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

tihs paper addresses the novelty attribution problem. This paper proposes Principal-Curve Novelty Geometry, a post-hoc measurement-contract framework that decomposes novelty in frozen sequential representations into progression, transverse deviation, and traversal rate. PCNG introduces an admissibility gate module and a principal-curve coordinate-and-rarity-calibration module.

Critical strengths
  1. The addresses problem is crucial in the frozen sequential representations field.
  2. The admissibility gate decides whether attribution is permitted, while the rarity calibration makes the three coordinates comparable once it is.
  3. The falsification apparatus is unusually complete, including a 3×3 confusion matrix and a residual scramble with a within-unit control.
  4. The unit-preserving randomisation stands on its own: PRONOSTIA’s coupling ratio falls from 7.1× to 1.17× once the reference respects units.
  5. the presentation is careful and well-structured.
  6. this framework can be used as a plug-and-play module applied to existing VLLMs.
Critical weaknesses
  1. The coordinate construction in the attribution module is a simple Hastie–Stuetzle principal curve followed by nearest-point projection, from which arc length, residual norm and its finite difference are read off and mapped to empirical-CDF tail probabilities.
  2. Appendix T returns a null on whether the coordinates answer different questions, and Table 32 shows U∥ at FA = 1.000 on 9/12 alarm jobs and Uṡ degenerate on 12/12. The three-coordinate frame supports one decision.
  3. The method introduces several sensitive hyperparameters, e.g., the Stage-1 loss weight λ_mono = 0.80, the 12 quantile bins used for residualisation, and the detector settings W = 20 and p98, none of which is varied.
  4. Lack of comparison with recent state-of-the-art methods, the prognostic-geometry methods identified as closest in intent — the polar (θ, r) representation of Thil et al. (2026) and the architectural-monotonicity approach of Zhao & Fink (2025).
  5. No external domain earns a GO on any seed, so the contract is never run end-to-end outside N-CMAPSS.
  6. Construct validity is measured on MonoCNN alone, which the authors trained with a monotone-degradation term, so the geometry the method needs is partly trained in.
  7. The headline 35/36 rests on 12 pairs × 3 seeds under one encoder. The paper rightly declines p-values, then lets the counts carry the weight anyway.
Questions for authors

Given Appendix T and Table 32, what is the operational case for keeping U∥ and Uṡ? Given Table 17 and the 0.149 label reproduction rate, what does the ternary label add over CV(L)? Can you report construct-validity interventions on one encoder you did not train? How is L_u computed when projections are non-monotone in arc length?

Reviewer_LVahOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a post-hoc measurement contract for a frozen sequential representation. A scalar novelty score does not say whether a window is unusual because of its position along a reference trajectory, its distance off that trajectory, or its speed along it. On training embeddings only, PCNG fits a principal curve (Hastie and Stuetzle, 1989, with a soft temporal-order term), projects each window onto it, and reads three coordinates: normalized arc length s, transverse distance d, and a within-unit finite-difference rate. Each is mapped to a reference-calibrated rarity. An admissibility gate uses the coefficient of variation of per-unit traversed length, with GO below 0.15, MARGINAL from 0.15 to 0.35, and NO-GO above 0.35. The gate and the coordinates come from the same training-only curve. No encoder weight is updated at this stage. The two encoders the authors train carry a monotone-degradation term; the other eleven configurations are frozen literature encoders. Stage 1 is explicitly not the contribution. The licensed reading is the transverse coordinate. Table 2 counts cells in which the intended component has the largest control-normalized response, out of 36 cells defined as 12 directed subset pairs times 3 seeds. A cross-subset failure-mode shift selects the transverse score in 35/36. A late-life holdout selects the progression score in 36/36, which the paper calls near-definitional because the ablation removes the arc-length range the probe occupies. An operating-condition holdout selects the rate score in 21/36, and the rate interpretation is refused. Cross-unit scrambling reduces the transverse hit rate from 0.972 to 0.667; a within-unit scramble stays at 1.000. The text says the cross-unit scramble cuts the hit count by roughly eleven cell-equivalents and that the 36 cells are not 36 independent trials, so no p-value is quoted. Residual coupling is reported as too heterogeneous for an independence claim. No external domain in Table 3 receives GO on any seed (PRONOSTIA and XJTU-SY are NO-GO on all three seeds; NASA Battery is MARGINAL on all three; NASA Capacitor straddles NO-GO). Unit bootstrap reproduces a run's ternary gate label in only 0.149 of resamples. Operationally, the abstract reports a 4.9% false-alarm rate for transverse rarity versus 29.2% for a learned early-degradation health monitor at matched coverage, and says distance-based novelty remains competitive. Appendix U finds that transverse rarity does not separate high-error windows from random selection.

Critical strengths

First, the contract refuses claims with counts, not with a paragraph of future work. Rate is 21/36 and is not licensed. No external domain earns GO. The progression row is labeled near-definitional. Selective prediction against random is reported as null in Appendix U. The unit-preserving randomization is the right null: it changes the PRONOSTIA coupling conclusion relative to window exchangeability, and the within-unit scramble moves one cell while the cross-unit scramble moves the hit rate from 0.972 to 0.667. That pattern is the part of the paper I would cite. Second, support artifacts are measured. The failure-mode ablation perturbs the progression coordinate, but the late-life holdout perturbs it more and selects the transverse score in 0/36. The paper also says the 36 cells are dependent and declines a p-value. A measurement paper that reports the negative control and the dependence is doing the job it set itself.

Critical weaknesses

First, the result that would have to carry acceptance is narrower than the abstract. Table 2's 35/36 is 12 directed subset pairs times 3 seeds, and the text says these are not independent trials. The abstract nests that count under 117 runs and 13 frozen configurations. I cannot tell from the abstract whether every configuration was subjected to the failure-mode intervention. A hit rate that falls only to 0.667, with the mass of the drop described as about eleven cell-equivalents, is a partial reduction. I would not recommend acceptance of a general attribution method for frozen sequential representations on this sample. Second, for a learning conference the new object is a protocol around a 1989 principal curve, and the paper does not compare that curve to a PCA line or to another principal-curve estimator. The encoders the authors train already include a monotone-degradation term, which is the progression structure the geometry is then asked to find. The operational comparison does not clear the alternative the authors themselves leave standing: distance-based novelty remains competitive, and the health-monitor gap is one operating point described as matched coverage. Gate labels that agree with a unit bootstrap only 0.149 of the time should not be read as stable GO or NO-GO decisions without the continuous CV(L), which the paper does report, but which the three-way label still invites readers to use.

Questions for authors

Clarification: Which of the 13 frozen configurations are included in the 36 cells of Table 2? If the answer is one encoder family, the abstract's sentence that joins 117 runs to 35 of 36 needs to be rewritten, and that rewrite would confirm the scope I am using for the rating. Clarification: Against a PCA-line baseline, and on a coverage sweep rather than a single matched-coverage point, does transverse rarity still separate from Mahalanobis distance and from the health coordinate? Section 8.2 already says matching Mahalanobis is compatible with the contribution. I need the magnitude of any remaining gap before I would raise the rating. Clarification: Where are the introduction's 1.19x and 1.1x representation-flatness figures computed? I could not trace them to a table. These are clarification requests about experiments the paper already frames. I am not asking for a new application domain as a condition of the current scores, beyond the encoder question, which changes what the existing table means.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.