Tag
This paper presents an audited protocol using orthogonal Procrustes alignment to detect donor-specific functional fingerprints in neural networks after Neural Collapse, demonstrating that distinct functional variations remain distinguishable even after convergence on MNIST with MLP-5 networks.