Online Learning via Learned Latent Bayesian Tracking
AURA learns a latent state space so deep models adapt online via Kalman filtering under distribution shift.
Researchers propose Adaptive Update through Representation Adaptation (AURA), which learns offline a low-dimensional latent model of how optimal parameters change under distribution shift. Online adaptation runs an extended Kalman filter in that space, then a lifting map rebuilds full model parameters in one step. On neural wireless receivers with time-varying channels and non-stationary image classification, AURA improves adaptation speed, accuracy, and compute versus online-learning and Bayesian-filtering baselines.
- AURA learns an offline low-dimensional latent model of parameter evolution.
- Online updates use extended Kalman filtering and a learned lifting map.
- Tested on time-varying wireless receivers and non-stationary image classification.
- Reports faster, more accurate, and cheaper adaptation than baselines.
From the original article
1 min read at arxiv.orgOnline learning in non-stationary environments requires models to adapt rapidly from streaming data under strict computational constraints. A principled approach casts online learning as Bayesian state tracking, where model parameters are updated sequentially via Bayesian filtering.
By Guy Gerson · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.