acceptodds
Under review as a conference paper at ICLR 2027

Path-Locked Neural Composition: Exact Evaluation-Path Invariance with Geometric Algebra

Abstract

We combine learned motion calibration with exact geometric algebra (GA) composition. Paired motor products in projective GA form a mergeable 16-coordinate summary, followed by a learned gate. On development-selected, family-held-out ETH3D odometry, this gate attains m translation error with 577 head parameters, compared with m for identity motion, a reduction. Its throughput is that of a contextual gated recurrent unit (GRU) variant at batch 256 on one CPU thread, excluding odometry extraction. Feature-matched direct and residual controls support the translation benefit of motor-constrained prediction. We characterize evaluation-path invariance by associativity on reachable states and evaluate it on four public recorded-data benchmarks. Fixed-input tests identify path-locking: MovieLens-1M accuracy falls from to when the same bilinear-residual model uses a right rather than left fold; TUM RGB-D motion error increases from to cm. Exact GA composition has a maximum observed path discrepancy of m on TUM. Intermediate-state regularization reduces empirical path sensitivity without certifying invariance. Associative motor composition supports accurate learned calibration and consistent chunk merging, while predictive improvements remain dataset- and horizon-dependent.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
geometric algebra, neural composition, associativity, evaluation-path invariance, mergeable representations, learned odometry calibration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_MoDqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. Problem and diagnostic. The paper studies “path-locking”: a learned binary composition operator trained with one reduction tree can change its internal state and prediction when the same ordered inputs are evaluated with a different parenthesization. It formalizes evaluation-path invariance as associativity over the reachable closure of encoded states and relates it to balanced reduction, chunk merging, and cached recomputation.

  2. Models. The paper compares unconstrained learned binary operators, mixed-path training, associator penalties, and exactly associative composition. For rigid motion, it represents SE(3) transforms as projective-geometric-algebra motors, equivalently unit dual quaternions, whose geometric product is associative. Its main ETH3D model pointwise-calibrates odometry increments, retains the raw and calibrated motor products as a 16-coordinate mergeable summary, and uses a small learned gate to blend them into a final motion.

  3. Evaluation. MovieLens and TUM provide fixed-input path-sensitivity tests. On MovieLens, changing a learned bilinear-residual model from a left to a right fold reduces accuracy from 0.4319 to 0.3483; the exact-matrix arm preserves every predicted rating across the tested trees, with only small floating-point logit differences. On TUM, a learned residual's median translation error changes from 6.83 to 36.53 cm, whereas motor composition agrees across tested trees up to numerical precision. On family-held-out ETH3D, the mergeable gate obtains 0.440 m translation error versus 0.915 m for identity motion and 0.526 m for a feature-matched direct head. A separately trained KITTI replication does not improve translation over raw odometry.

Critical strengths
  1. A clean structural guarantee with a genuinely mergeable summary. In exact arithmetic, the paired raw/calibrated motor state composes associatively, preserves motion order, and supports chunk-wise reduction and cache updates without a recurrent encoder. The paper correctly distinguishes merging this complete summary from merging independently gated predictions and does not claim that PGA is more accurate than equivalent dual-quaternion or homogeneous-matrix coordinates.

  2. Unusually careful and transparent experimentation. Inputs and checkpoints are held fixed in the path tests; uncertainty is clustered by users, trajectories, or families rather than overlapping windows. The paper discloses development selection, post-hoc controls, dependent folds, an unopened Oxford confirmation set, a failed KITTI utility criterion, and unfavorable rotation and long-horizon results. The artifact reportedly includes deterministic splits, checksums, checkpoints, complete run records, and exactness audits.

Critical weaknesses
  1. The central formal and diagnostic claims have limited novelty and an unclear target. Theorem 1 is the classical equivalence between associativity and parenthesization independence, and Proposition 1 gives standard parallel-reduction and segment-tree consequences. Exact-arithmetic matrix or motor composition being invariant is true by construction. Conversely, the tested learned operators are not designed to be associative, so some change under a new tree is unsurprising. The useful result is the measured magnitude of the change, not the existence of “path-locking.” Moreover, the headline MovieLens right-fold test changes the masked query's depth as well as the history grouping; the query-last control still shows a drop, but the bilinear effect is smaller, from 0.4319 to 0.4032 rather than 0.3483. The paper needs a clearer demonstration that arbitrary regrouping is a realistic requirement for the learned models being diagnosed.

  2. The predictive evidence for the proposed motor-calibration architecture is narrow and not independently confirmatory. ETH3D was inspected through several earlier protocols before the family-held-out experiment, and the strongest feature-matched controls were post-hoc. The evaluation has only 16 named families across folds with overlapping training sets; conditional family bootstraps omit retraining uncertainty, while the fold-block sensitivity for mergeable versus feature-direct gives p=0.0625p=0.0625. The gain is also translation-specific: feature-matched direct and residual heads have lower rotation error than the mergeable gate (24.81/24.65 versus 28.60 degrees). Most importantly, the separately specified KITTI replication fails its predictive criterion: mergeable translation error is 1.168 m versus 1.062 m for raw stereo odometry. Thus, associativity is established, but robust calibration accuracy is not.

  3. The experiments do not establish a distinctive benefit from geometric algebra. PGA motor multiplication here is equivalent to unit-dual-quaternion multiplication and to rigid-transform composition under a coordinate conversion, as the paper acknowledges. On TUM, the exact leaves are relative transforms derived from tracked poses and the target is their endpoint composition, so near-zero exact-product error primarily validates implementation and conventions. On ETH3D, the gain can arise from pointwise calibration and the learned raw/calibrated blending prior; the controls do not attribute it to associativity or PGA. There is also no comparison with a standard odometry-correction or filtering method under the same motion-only inputs, and the negative KITTI result shows that the learned correction can damage an already accurate front end.

  4. The efficiency result has limited practical scope. The reported 2.17× gain is batch-256, single-thread CPU throughput and excludes RGB-D odometry extraction. At batch one, latency changes only from 2.360 to 2.158 ms. No end-to-end SLAM speedup or measured parallel-tree/cache-update speedup is demonstrated, so the systems claim should remain about algebraic mergeability and removal of the GRU rather than deployment efficiency.

Questions for authors
  1. Which deployed model classes are intended to permit arbitrary reassociation? For a learned operator specified as a left recurrence, why should a right-fold computation be considered another evaluation path of the same model rather than a different computation graph? Can the primary diagnostic keep the query depth fixed prospectively?

  2. Beyond applying classical associativity to a behavioral test, what is the precise novel technical contribution? Since equivalent dual-quaternion and homogeneous-transform implementations have the same merge law, which claims specifically require projective geometric algebra?

  3. Can the authors construct and evaluate a fully sign-invariant gate and readout whose output depends only on the represented SE(3) motions, rather than convention-fixed motor coordinates? The Appendix L alignment is applied after the potentially sign-sensitive gate.

  4. Can the complete frozen ETH3D architecture be evaluated on genuinely untouched environments or another dataset, with no model or threshold reselection? Please include retraining-aware uncertainty, standard motion-only calibration/filtering baselines, and translation and rotation criteria jointly.

  5. What practical workload benefits from arbitrary chunk regrouping in this application? Please report end-to-end or cache-update latency, including feature or odometry extraction where relevant, rather than only batch-256 aggregation throughput.

Reviewer_edzeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies prediction changes caused by reparenthesizing an ordered sequence while keeping inputs and checkpoints fixed. It combines pointwise motion calibration with paired exact PGA motor products and a learned gate, yielding a mergeable 16-coordinate summary. Experiments diagnose path sensitivity on MovieLens and TUM and evaluate learned odometry correction on ETH3D and KITTI. The main predictive improvement is on ETH3D; KITTI does not show improved primary translation accuracy.

Critical strengths
  1. The fixed-input diagnostic separates trained-fold accuracy from sensitivity to reparenthesization. Five-seed experiments, clustered uncertainty estimates, and query-last controls provide convincing evidence of path sensitivity in the tested learned composers.
  2. The paired motor summary combines learned calibration with exact ordered chunk merging. On ETH3D, the compact gate reduces translation error from 0.526 to 0.440 m relative to feature-matched direct regression, with transparent reporting of limitations.
Critical weaknesses
  1. The methodological novelty is limited. The invariance and merge/cache guarantees follow from classical associativity. The paper does not sufficiently establish an advance beyond combining existing exact composition, learned calibration, and endpoint blending.
  2. Predictive gains remain dataset-dependent: KITTI translation error worsens from 1.062 to 1.168 m, and ETH3D direct/residual heads retain lower rotation error. The reported throughput gain reflects GRU removal, leaving the practical benefit of associative merging insufficiently isolated.
Questions for authors
  1. What new capability does the method provide beyond established exact rigid-motion composition, calibration, and blending? Can this be demonstrated in a practical chunk-merging or cache-update task?
  2. Can you validate the predictive benefit on independent data and compare against equally compact endpoint-only direct/residual heads, reporting both pose accuracy and measured merge/update latency?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.