In independently trained neural networks, the lack of a shared neuron-index reference frame necessitates accounting for coordinate freedom during comparisons. Neural Collapse exacerbates this issue: networks converge toward a shared, low-dimensional geometry, raising the question of whether trajectory-specific functional variation remains distinguishable post-convergence. This study distinguishes three claims: detectability, transplantability, and causal persistence, focusing primarily on detectability.
Using five independently trained networks reconstructing Neural Collapse on the MNIST dataset, a verified affine-correct alignment mapping was applied to align donor heads into recipient coordinates. After baseline correction at the recipient level, donor-specific functional fingerprints remained distinguishable: all 20 ordered donor-recipient pairs were correctly identified, with an exact permutation p=0.0083, robust against a leakage audit. These findings establish detectability under the test employed, but do not confirm transplantability or causal persistence.
The study illustrates how alignment, ambiguity diagnostics, and leakage control can be combined to test cross-network variation in a controlled setting; whether this can generalize beyond it remains an open question.
Blogger's Review: This paper presents a novel auditing protocol in neural network comparison, emphasizing the detectability of donor features post-Neural Collapse. Future research should explore the transplantability and causal relations of these findings to advance the field further.