aiqbench analysis
This is a theoretical machine learning paper proposing spectral diagnostics (a "Global Matrix Stability Index" and spectral entropy) for how sensitive deep networks are to input perturbations, label noise, and training dynamics. Its empirical claims are modest, limited to MNIST and CIFAR experiments showing that spectral regularization improves attribution stability. The accompanying note flags two theorems (4.2 and 5.1) as incorrect or incomplete, which undercuts the central theoretical contribution, so practitioners should treat the framework as unverified until those proofs are corrected.
A Unified Matrix-Spectral Framework for Stability and Interpretability in Deep Learning
arXiv:2602.01136v2 Announce Type: replace Abstract: We develop a unified matrix-spectral framework for analyzing stability and interpretability in deep neural networks. Representing networks as data-dependent products of linear operators reveals spectral quantities governing sensitivity to input perturbations, label noise, and training dynamics. We introduce a Global Matrix Stability Index that aggregates spectral information from Jacobians, parameter gradients, Neural Tangent Kernel operators, and loss Hessians into a single stability scale controlling forward sensitivity, attribution robustness, and optimization conditioning. We further show that spectral entropy refines classical operator-norm bounds by capturing typical, rather than purely worst-case, sensitivity. These quantities yield computable diagnostics and stability-oriented regularization principles. Synthetic experiments and controlled studies on MNIST, CIFAR-10, and CIFAR-100 confirm that modest spectral regularization substantially improves attribution stability even when global spectral summaries change little. The results establish a precise connection between spectral concentration and analytic stability, providing practical guidance for robustness-aware model design and training.
Abstract:We develop a unified matrix-spectral framework for analyzing stability and interpretability in deep neural networks. Representing networks as data-dependent products of linear operators reveals spectral quantities governing sensitivity to input perturbations, label noise, and training dynamics. We introduce a Global Matrix Stability Index that aggregates spectral information from Jacobians, parameter gradients, Neural Tangent Kernel operators, and loss Hessians into a single stability scale controlling forward sensitivity, attribution robustness, and optimization conditioning. We further show that spectral entropy refines classical operator-norm bounds by capturing typical, rather than purely worst-case, sensitivity. These quantities yield computable diagnostics and stability-oriented regularization principles. Synthetic experiments and controlled studies on MNIST, CIFAR-10, and CIFAR-100 confirm that modest spectral regularization substantially improves attribution stability even when global spectral summaries change little. The results establish a precise connection between spectral concentration and analytic stability, providing practical guidance for robustness-aware model design and training.
Theorem 4.2 incorrectly claims that an upper spectral scale controls NTK conditioning without a lower eigenvalue bound. Theorem 5.1 also omits the Jacobian's magnitude from its spectral-entropy sensitivity bound. These problems undermine the central theoretical contribution
Machine Learning (cs.LG); Dynamical Systems (math.DS); Optimization and Control (math.OC)
arXiv:2602.01136 [cs.LG]
(or arXiv:2602.01136v2 [cs.LG] for this version)
Full text under the publisher's open licence · 199 words