IC-917ViT-S models show early-layer sensitivity to layer-wise averaging, with the averaging direction being far more disruptive than random perturbations of the same norm

Linara Adilova, Maksym Andriushchenko, Michael Kamp, Asja Fischer, Martin Jaggi

SourceLayer-wise linear mode connectivity

The paper evaluates three pairs of pre-trained ViT-S models (differing in learning rate, data augmentation, or SAM radius) from a public repository. Layer-wise interpolation at alpha=0.5 is barrier-free for almost all layers, but the most sensitive layers are the early attention (to-qkv) and fully-connected (net) weights, with barriers reaching ~4-5 in early layers versus ~0 in deep layers. When comparing the loss increase from layer-wise interpolation to random perturbations of the same norm, the interpolation direction causes substantially higher loss, indicating that the averaging direction is a special, high-curvature direction on the loss surface. Flatter models (larger learning rate or larger SAM rho) show slower loss growth over the interpolation coefficient.

Evidence
correlational
Key metric
to-qkv layer barrier up to ~5.3 (small lr) vs. ~0.0 for deep layers; random-direction perturbation loss ~0.0-0.1 for the same layers
Caveat
Results are for ViT-S on CIFAR-10 only. The paper notes that the success of interpolations is highly asymmetric for a pair of models.
Model
ViT ViT-S
Concepts
Depth-dependent structure
Datasets
CIFAR-10 [eval]
Related findings
IC-916
Extraction
automatic-extraction