Beyond Coordinate Gauge: An Audited Protocol for Detecting Donor-Specific Functional Fingerprints after Neural Collapse

2026-07-12 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Independently trained neural networks pose a comparison challenge due to their lack of a shared neuron-index reference frame, a problem intensified by Neural Collapse where networks converge to a shared low-dimensional geometry. This study investigates whether trajectory-specific functional variation remains distinguishable after such convergence, focusing on "detectability." Researchers applied an affine-correct alignment mapping to five independently trained networks reconstructing Neural Collapse on MNIST. They found that donor-specific functional fingerprints remained distinguishable even after recipient-level baseline correction. All 20 ordered donor-recipient pairs were correctly identified, achieving an exact permutation p=0.0083, robust to a leakage audit. These findings establish detectability under the specific test used, demonstrating a protocol combining alignment, ambiguity diagnostics, and leakage control for cross-network variation analysis, though generalizability beyond this setting is an open question.

Key takeaway

For AI Scientists or Machine Learning Engineers analyzing independently trained neural networks, this research suggests that even after Neural Collapse, you can detect donor-specific functional fingerprints. If you are developing methods for cross-network comparison or transfer learning, consider incorporating affine-correct alignment and baseline correction. This approach provides a robust protocol for identifying subtle functional variations, which is crucial for understanding network behavior and potentially improving model interpretability or robustness.

Key insights

Donor-specific functional fingerprints are detectable in independently trained neural networks after Neural Collapse using affine alignment and baseline correction.

Principles

• Independently trained networks lack shared reference frames.
• Neural Collapse complicates cross-network comparison.
• Affine-correct alignment enables cross-network mapping.

Method

Apply affine-correct alignment mapping to donor heads into recipient coordinates. Then, perform recipient-level baseline correction to distinguish donor-specific functional fingerprints.

Topics

• Neural Collapse
• Network Alignment
• Functional Fingerprints
• Cross-Network Comparison
• MNIST Dataset
• Deep Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability — Tensor similarity offers a weight-based, symmetry-invariant metric for comparing neural network components functionally.
• Decodable but Not Detectable: A Leakage Fingerprint for Near-OOD Benchmarks — Benchmark leaks in near-OOD evaluation can invert detector scores, requiring a specific diagnostic to identify contamination.
• Curvature-Aligned Probing for Local Loss-Landscape Stabilization — Curvature-aligned probing in top Hessian eigenspaces offers a more efficient and accurate way to assess loss-landscape stabilization.
• Decoupled Alignment for Robust Plug-and-Play Adaptation — Dapa efficiently aligns LLMs against harmful content by surgically transferring alignment knowledge via memory editing.
• Evaluating Fine-Tuning and Metrics for Neural Decompilation of Dart AOT Binaries — Neural decompilation fine-tuning often fails to improve functional correctness, and superficial metrics can mislead evaluation.
• Attention as Frustrated Synchronization — Attention can be modeled as frustrated synchronization, outperforming transformers at similar scales.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.