PEIRA · Results

The predicted alignment on real images (CIFAR-10)

One PEIRA training run on CIFAR-10: signal and noise align their eigenvectors, new spectral modes emerge, and the objective tracks accuracy.

Dynamics of PEIRA on CIFAR-10 Three plots drawn from per-epoch data of one training run (ResNet-18, CIFAR-10, 1000 epochs), shown one at a time. First: alignment of the top eigenvectors of the noise matrix with the signal matrix rises toward 1 over epochs, one curve per 100 epochs. Second: the spectrum of the noise matrix develops additional significant modes. Third: validation top-1 (reaching 91.93%) and the negative PEIRA objective rise together over training. One PEIRA training run on CIFAR-10: signal and noise align their eigenvectors. PANEL 1 OF 3 Alignment rises toward 1 Top eigenvectors of N become eigenvectors of Σ. α_{i} = e_{i}^{⊤}Σ_{U,V}e_{i} / (‖e_{i}‖ ‖Σ_{U,V}e_{i}‖) ei: i-th eigenvector of the noise NU,V αi = 1: ei is also an eigenvector of ΣU,V PANEL 2 OF 3 More modes emerge The spectrum of N gains significant directions. spectrum of N_{U,V} Log scale; one curve per 100 epochs, from early (purple) to late (yellow). PANEL 3 OF 3 Objective tracks accuracy Linear-probe top-1 rises with the negative PEIRA objective. −ℰ_{PEIRA} vs top-1 Strongly correlated during training: a possible label-free quality indicator. DATA One training run (ResNet-18, CIFAR-10, 1000 epochs, λ = 0.7, seed 42). Real per-epoch data from one run: one spectral curve per 100 epochs; accuracy and objective every 5 epochs.

0:00 / 0:15