Online Learning via Learned Latent Bayesian Tracking
AURA turns Bayesian online learning into latent-state tracking instead of filtering full deep-model parameters.
The paper frames high-dimensional Bayesian filtering as the bottleneck for adapting deep models to streaming distribution shifts. Its proposed meta-learning method learns a low-dimensional latent dynamics offline, then uses extended Kalman filtering there during online updates. Full model parameters are reconstructed through a learned lifting map, aiming for single-step adaptation without giving up model expressiveness. The authors report gains in speed, accuracy, and compute efficiency on neural wireless receivers and non-stationary image classification. ArXiv · AI/CL/LG's note
The paper frames high-dimensional Bayesian filtering as the bottleneck for adapting deep models to streaming distribution shifts. Its proposed meta-learning method learns a low-dimensional latent dynamics offline, then uses extended Kalman filtering there during online updates. Full model parameters are reconstructed through a learned lifting map, aiming for single-step adaptation without giving up model expressiveness. The authors report gains in speed, accuracy, and compute efficiency on neural wireless receivers and non-stationary image classification. ArXiv · AI/CL/LG's note
score 4