Online Learning via Learned Latent Bayesian Tracking
Abstract Overview
The paper studies online learning in non-stationary environments where models must adapt from streaming labeled samples under tight computational constraints. It argues that direct Bayesian filtering in the full parameter space of deep models is impractical, and proposes AURA, which performs adaptation in a learned low-dimensional latent state space while reconstructing full model parameters through a learned lifting map. Online updates are carried out with a single-step extended Kalman filter in latent space, and the latent dynamics, noise parameters, and lifting map are meta-learned offline from non-stationary trajectories. The method is evaluated on non-stationary image classification and neural wireless receiver adaptation, with additional discussion of complexity-performance trade-offs and deployment considerations.
Novelty
The distinctive contribution is to frame the main bottleneck in Bayesian online learning as the lack of a learned low-dimensional dynamical representation, rather than the filter itself. AURA meta-learns that latent state-space representation and uses EKF updates in the latent space, avoiding manually specified adaptation subspaces or restrictive parameter-space covariance approximations.
Results
On gradual non-stationary image classification, AURA achieves the best reported error among streaming-feasible methods and is best overall on MNIST-C and CIFAR-10-C, while ranking second overall on CIFAR-100-C. In wireless receiver adaptation, it consistently attains the lowest BER across both linear and nonlinear channel settings and shows more stable temporal tracking than the compared gradient-based and parameter-space Bayesian baselines. The paper also reports that latent-space adaptation gives a more favorable complexity-accuracy trade-off than parameter-space filtering.
Key Points
- AURA performs single-step online adaptation by tracking a low-dimensional latent state with an extended Kalman filter and then lifting that state back to full model parameters.
- The latent dynamics, lifting map, and optionally the noise covariances are learned offline through trajectory-based meta-learning optimized for online adaptation performance under distribution shift.
- Experiments across image corruption benchmarks and wireless communication tasks support that a learned adaptation-aware latent geometry improves both efficiency and adaptation quality relative to streaming baselines.