Canonical locks that encode part-whole hierarchies
Rajat Modi, Yogesh Singh Rawat
cs.AI
Quick verdict
Proposes canonical locks: part-whole hierarchies encoded as relative phase angles in GLOM-style nets. Rich concepts; empirical evidence remains preliminary.
Summary
This position/method paper proposes 'canonical locks' as a geometric primitive for encoding part-whole hierarchies in neural networks, advancing a second generation of Asynchronous Perception Machines (APMs) toward Hinton's GLOM program. Parts and wholes are modeled as high-dimensional vectors (d≥4); their relation is carried by a constant relative phase angle that persists while both vectors rotate. The architecture pairs a shared encoder with per-level bottom-up and top-down nets that iterate toward agreement: part votes are projected onto whole vectors, a geodesic 'slurp' update nudges parts, and a label-free Deep Weight Alignment loss drives the top-down net toward the bottom-up inverse, yielding signed symmetry in weight visualizations. Toy experiments show symmetry breaking in ~30 iterations for anti-symmetric configurations versus ~60 for collapsed ones, and a two-point linear fit predicts ~22.5 iterations at a 225-degree offset, reportedly confirmed. On CelebA-HQ the model parses faces into a hand-defined 5-level hierarchy, decoding paths directly from classifier heads. Key limitations: headline claims of one-sample training with 'competitive accuracies' lack baseline tables; the linear law rests on only two fitted points; experiments are toy-scale; no code is released; and quantitative findings are explicitly deferred to a future paper.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Part-whole relations can be encoded as canonical locks: constant relative phase angles between high-dimensional part and whole vectors that persist while both rotate (§3, §6.1).
- A label-free inner loop (agreement procedure + slurp update + Deep Weight Alignment) drives the top-down net toward the inverse of the bottom-up net, producing signed weight symmetry (§7.2-7.5, Fig 30).
- The iterations needed to break a symmetric part-whole configuration depend linearly on the center-of-gravity angle: 30 iterations at 180°, 60 at 0°, predicted and observed ~22.5 at 225° (§5.3).
- APM-2 trains on a single sample without pretraining, generalizes to a full test set with competitive accuracy, and densely parses CelebA-HQ faces into a hand-defined 5-level hierarchy readable directly from classifier outputs (§8-9, Appendix A).
Strengths
Concrete GLOM-to-engineering translation
The agreement procedure (§7.2, Eqs 1-6), geodesic slurp update (§7.3), and Deep Weight Alignment loss (§7.4, Eq 8) turn GLOM's philosophy into an implementable, label-free inner loop (Algorithm 1) — rare specificity in a literature where, as the authors note, a GLOM implementation 'has continued to remain elusive'.
Falsifiable symmetry-breaking phenomenology
§5 isolates anti-symmetric (90/180/270°) and collapsed (0°) part-whole configurations and measures breaking time (~30 vs ~60 iterations, Figs 8-9); the two-point linear law predicts 22.5 iterations at 225°, and Fig 11 shows the crossing — a novel, checkable dynamic.
Candid self-assessment
Post-hoc-free hierarchical decoding
Weaknesses
Headline empirical claim unsupported
Two-point linear law
Toy-scale evidence
Analogy-driven argumentation and an internal inconsistency
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This preprint (arXiv 2609.26046, cs.AI, dated Sept 23, 2026) is a concept/position paper introducing **canonical locks**: a geometric primitive in which parts and wholes are high-dimensional vectors (d≥4) whose relation is a constant relative phase angle, maintained while both vectors rotate forever (the 'dynamical system' stance, §3.2). It presents this as a second generation of the authors' Asynchronous Perception Machine (APM, Modi & Rawat 2024), explicitly framed as translating Hinton's GLOM philosophy into 'engineering constraints' (§1). The authors are candid that quantitative findings are deferred: 'A subsequent paper (currently under drafting) shall present our quantitative findings' (§1). Central claims: (i) canonical locks encode part-whole hierarchies without N×M routing; (ii) a label-free inner loop (agreement procedure §7.2, slurp update §7.3, Deep Weight Alignment §7.4) drives top-down weights toward the bottom-up inverse, yielding signed symmetry (Fig 30); (iii) symmetry-breaking time depends linearly on the part-whole center-of-gravity angle (§5.3); (iv) APM-2 trains on one sample, generalizes with 'competitive accuracies', and densely parses CelebA-HQ faces into a hand-built 5-level hierarchy (11/7/5/3/2 classes, §8) read directly from classifier heads (Appendix A).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is a 'canonical lock' in this paper?
A learned configuration where a part vector and its whole vector maintain a constant relative angle (phase offset) even as both vectors keep rotating over time. The invariant relative angle — not absolute direction or magnitude — is what encodes the part-whole relationship.
Is this paper an implementation of Hinton's GLOM?
It is a partial, interpretation-level step toward GLOM: it defines bottom-up and top-down neural fields that reach 'islands of agreement' without routing. The authors explicitly state that full quantitative results will appear in a subsequent paper, and no comparison against other methods is provided here.
Does the paper show state-of-the-art results?
No. The abstract claims competitive accuracies when training on a single sample, but the text contains no accuracy tables or baselines. The only numbers given are per-level accuracies (0.919 to 0.982) for a single CelebA-HQ face in the appendix.
What data and experiments does the paper include?
Toy symmetry-breaking experiments on synthetic configurations of four vectors (anti-symmetric configurations break in ~30 iterations, collapsed ones in ~60), plus visual parsing of CelebA-HQ faces into a hand-defined 5-level hierarchy with 11/7/5/3/2 classes per level.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are available only as captions/descriptions in the extracted text; the cosine-similarity curves (Figs 8, 9, 11), the Hinton diagram (Fig 30), and the parsing output (Fig 1) could not be visually verified.
- No accuracy tables for the headline one-sample-generalization claim appear anywhere in the extracted text; they may exist only inside figures.
- The Appendix A JSON dump appears truncated at the end of the extraction.
- Forward-dated citations (2026) could not be verified against the literature.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.