When neural networks finish training, do they all end up looking the same? A new study from researchers including Truong Xuan Khanh challenges that assumption — and the answer turns out to be more nuanced than either side of the debate might expect. The research tackles a problem at the heart of modern machine learning: detecting neural fingerprints that survive a powerful convergence phenomenon, even when networks trained independently have no shared reference frame to begin with.
Summary
Key takeaways
- Neural networks trained independently have no shared coordinate system, requiring alignment before meaningful comparison is possible.
- Neural Collapse pushes networks toward a shared low-dimensional geometry, but donor-specific functional fingerprints remain detectable afterward.
- Using five independently trained networks on MNIST, all 20 ordered donor-recipient pairs were correctly identified with a permutation p-value of 0.0083.
- Results held under a leakage audit, confirming methodological rigor.
- The study establishes detectability only — transplantability and causal persistence of these fingerprints remain open questions.
Neural Collapse and Coordinate Freedom in Network Comparison
Comparing two independently trained neural networks is harder than it sounds. Each network develops its own internal coordinate system — there is no shared neuron-index reference frame across models. Before any meaningful comparison can happen, researchers must account for this coordinate freedom, essentially solving an alignment problem before even asking what differences exist.
Challenges in Comparing Independently Trained Networks
This problem is not new, but a specific training phenomenon called Neural Collapse sharpens it considerably. As networks approach convergence during training, their learned representations tend to compress toward a shared, low-dimensional geometry. The last layers of the network reorganize into tight, symmetric structures that look strikingly similar across independently trained models.
That convergence raises a genuinely uncomfortable question for researchers: if networks settle into roughly the same geometric shape, does anything distinctly individual survive? Or does Neural Collapse wash out the functional differences that arose during each network’s unique training trajectory?
Shared Low-Dimensional Geometry Post Neural Collapse
The answer, according to this research, is that something does survive — but detecting it requires very careful methodology. The study frames the problem around three distinct concepts: detectability, transplantability, and causal persistence. These are not the same thing, and conflating them has muddied previous discussions in the field. The researchers focus exclusively on detectability, which is the most tractable of the three and the logical first step.
Experimental Protocol for Detecting Donor-Specific Fingerprints
The experimental design is deliberately controlled and auditable. Five independently trained networks were used to reconstruct Neural Collapse on the MNIST dataset — a well-known benchmark of handwritten digit classification. From these five networks, the researchers constructed all possible ordered donor-recipient pairs, yielding 20 combinations to test.
Using Five Independently Trained Networks on MNIST Dataset
The choice of MNIST provides a clean, low-noise testing environment. Each network trained on the same data but independently, meaning any detectable differences between them reflect divergence in their training trajectories rather than data artifacts. This controlled setup is important: it allows the researchers to isolate the signal they are looking for without confounding variables from dataset variation.
Affine-Correct Alignment Mapping Methodology
The methodological centerpiece of the study is an affine-correct alignment mapping that transforms each donor network’s internal representations into the coordinate system of the recipient network. This step is non-trivial. Without proper alignment, comparing functional patterns across networks is essentially comparing measurements taken in different units — the numbers may look different simply because the rulers are different.
After alignment, the researchers applied a recipient-level baseline correction. This strips out variation that comes from the recipient network itself, leaving only what is genuinely attributable to the donor. The combination of affine alignment and baseline correction is what makes the detection approach rigorous rather than speculative.
Results Confirm Detectability of Functional Fingerprints
The results are clear-cut within the scope of the experiment. Donor-specific functional fingerprints remained distinguishable even after baseline correction — meaning the individual identity of each donor network left a measurable trace that could be reliably separated from background variation.
Distinguishability after Baseline Correction
The strength of this finding lies in how clean the discrimination turned out to be. Across all 20 ordered donor-recipient pairs, every single pairing was correctly identified. There were no misclassifications, no ambiguous cases. That is a perfect classification result across the full set of combinations derived from five networks.
Statistical Significance and Robustness through Leakage Audit
The statistical significance of that outcome was assessed using an exact permutation test, yielding a p-value of 0.0083. This is well below conventional thresholds for significance and indicates the result is extremely unlikely to be a product of chance given the experimental design.
Critically, the findings held up under a leakage audit — a methodological check designed to detect whether information from the donor inappropriately bled into the baseline correction process. The audit finding matters: it rules out the possibility that the apparent detectability was an artifact of how the experiment was set up, rather than a genuine property of the networks themselves. In machine learning research, where overfitting and data leakage regularly undermine seemingly strong results, passing a leakage audit is a meaningful form of validation.
Limitations and Open Questions
The study is deliberate about what it does and does not claim. Detectability is established under the specific conditions tested here. Transplantability — whether a donor fingerprint could be meaningfully transferred into a recipient network — and causal persistence — whether these fingerprints actually cause observable behavioral differences — remain entirely unverified. The researchers do not speculate beyond their evidence.
That epistemic restraint is notable. The broader machine learning field frequently conflates detectability with deeper claims about identity or causation. By explicitly distinguishing the three concepts and addressing only the first, this work sets a higher methodological bar for follow-up research. Whether the approach scales beyond a controlled MNIST experiment — to larger datasets, more complex architectures, or real-world deployment contexts — is an open question the study acknowledges directly.
The work demonstrates how alignment, ambiguity diagnostics, and leakage control can be combined into a testable protocol for studying cross-network variation. That framework itself may be as significant as the specific findings: it provides a replicable structure that future research can stress-test against harder problems. The deeper puzzle — whether these fingerprints mean anything functionally beyond their detectability — remains unsolved.
FAQ
What is Neural Collapse and why does it matter in this study?
Neural Collapse is the phenomenon where networks converge toward a shared low-dimensional geometry during training. It matters here because it raises the question of whether individual functional variation between networks survives that convergence — and whether any remaining differences are still detectable.
How did the researchers detect donor-specific functional fingerprints after convergence?
They applied an affine-correct alignment mapping to transform donor networks into the coordinate system of a recipient network, then applied recipient-level baseline correction. This process isolated donor-specific patterns from background variation, allowing successful identification of fingerprints.
What were the main findings regarding the detectability of donor-specific fingerprints?
All 20 ordered donor-recipient pairs derived from five independently trained networks were correctly identified, with an exact permutation p-value of 0.0083. The results were also robust to a leakage audit, confirming the methodological soundness of the detection approach.
Does the study confirm that these fingerprints can be transplanted or persist causally?
No. The study confirms detectability only. Whether donor fingerprints can be transplanted into recipient networks or whether they causally drive observable behavioral differences remains unverified and outside the scope of this research.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
