Boosting Data Augmentation with Stochastic Weight Averaging
SWA is presented as a cheaper route to the symmetry gains seen in large augmented-data ensembles.
The paper studies stochastic weight averaging as an alternative to repeatedly training deep ensembles. Its analysis models the late training path as an Ornstein-Uhlenbeck process and argues that, at infinite width, SWA with augmentation improves equivariance beyond SWA’s ordinary performance lift. The authors report experiments across vision and graph classification tasks, covering discrete and continuous symmetries. ArXiv · AI/CL/LG's note
The paper studies stochastic weight averaging as an alternative to repeatedly training deep ensembles. Its analysis models the late training path as an Ornstein-Uhlenbeck process and argues that, at infinite width, SWA with augmentation improves equivariance beyond SWA’s ordinary performance lift. The authors report experiments across vision and graph classification tasks, covering discrete and continuous symmetries. ArXiv · AI/CL/LG's note
score 4