Canonicalization Leakage: How Canonical Representatives Confound Supervised Learning under Group Symmetry

Fuente: Zenodo
Saved in:
Bibliographic Details
Main Author: Li, Alex
Format: Recurso digital
Language:English
Published: Zenodo 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866901925725732864
author Li, Alex
author_facet Li, Alex
contents <p>Models trained on canonical representatives of equivalence classes under group symmetry can exploit representation artifacts rather than learning invariant structure. We propose CL-DIAG, a six-step diagnostic protocol that detects, localizes, and quantifies this "canonicalization leakage." Applied to circuit complexity prediction over 616,126 NPN equivalence classes of 5-input Boolean functions (|G| = 7,680), CL-DIAG reveals that a baseline MLP achieves Spearman r_s = 0.788 on canonical data but only r_s = 0.254 when NPN-averaged, with 0% prediction consistency. Signal decomposition shows canonical performance decomposes into classical invariant signal (r_s = 0.635), neural invariant signal (+0.142), and canonicalization leakage (+0.011). NPN augmentation at 7x recovers r_s = 0.777, exceeding the classical invariant ceiling by 14 percentage points. A matched-volume control confirms the gain is from symmetry-consistent augmentation, not generic regularization.</p> <p>v2: Figure 1 now uses actual model predictions (previously used placeholder visualization). No changes to text, results, or conclusions.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19112504
institution Zenodo
language eng
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Canonicalization Leakage: How Canonical Representatives Confound Supervised Learning under Group Symmetry
Li, Alex
canonicalization leakage
NPN equivalence
group symmetry
data augmentation
shortcut learning
Boolean functions
circuit complexity
<p>Models trained on canonical representatives of equivalence classes under group symmetry can exploit representation artifacts rather than learning invariant structure. We propose CL-DIAG, a six-step diagnostic protocol that detects, localizes, and quantifies this "canonicalization leakage." Applied to circuit complexity prediction over 616,126 NPN equivalence classes of 5-input Boolean functions (|G| = 7,680), CL-DIAG reveals that a baseline MLP achieves Spearman r_s = 0.788 on canonical data but only r_s = 0.254 when NPN-averaged, with 0% prediction consistency. Signal decomposition shows canonical performance decomposes into classical invariant signal (r_s = 0.635), neural invariant signal (+0.142), and canonicalization leakage (+0.011). NPN augmentation at 7x recovers r_s = 0.777, exceeding the classical invariant ceiling by 14 percentage points. A matched-volume control confirms the gain is from symmetry-consistent augmentation, not generic regularization.</p> <p>v2: Figure 1 now uses actual model predictions (previously used placeholder visualization). No changes to text, results, or conclusions.</p>
title Canonicalization Leakage: How Canonical Representatives Confound Supervised Learning under Group Symmetry
topic canonicalization leakage
NPN equivalence
group symmetry
data augmentation
shortcut learning
Boolean functions
circuit complexity
url https://doi.org/10.5281/zenodo.19112504