Efficient online learning through learned low dimensional parameter tracking
Online Learning via Learned Latent Bayesian Tracking
Machine Learning
Summary
Online learning means updating a model quickly as new information comes in, but it is very hard to do this with big deep learning models because they have too many parts to adjust at once. The authors found that the problem is not having a simple enough way to represent how the best model parameters change over time. They created a system called AURA that learns a smaller, simpler map of these changes ahead of time, so it can update the full model quickly when new data arrives. This helps models adapt faster and with less computation in changing environments like wireless signals or shifting image categories.
What this means in practice
- •For wireless communication engineers: Adapt wireless receivers rapidly to changing channel conditions using efficient latent online learning.$Commercial implications: Enables faster and more accurate adaptation in communication devices, improving performance in dynamic wireless environments.
- •For machine learning engineers: Improve model adaptation speed and accuracy for non-stationary image classification tasks with low computational cost.
Authors
Guy Gerson, Tomer Raviv, Nir Shlezinger, Tirza Routtenberg, Osvaldo Simeone
Abstract
Online 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. However, applying Bayesian filters directly to modern deep models is computationally prohibitive due to the high dimensionality of parameter space, forcing existing methods to rely on restrictive approximations or manually designed low-dimensional subspaces. In this work, we identify the absence of a suitable low-dimensional dynamical representation as the core bottleneck in Bayesian filtering-based online learning. Accordingly, we propose Adaptive Update through Representation Adaptation (AURA), a meta-learning framework that learns offline a low-dimensional latent state-space model governing the evolution of optimal model parameters under distribution shift. Online adaptation is then performed via extended Kalman filtering in this learned latent space followed by reconstruction of the full model parameters through a learned lifting map, enabling efficient single-step online adaptation while preserving model expressiveness. Evaluated on online adaptation of neural wireless receivers under time-varying channels and on non-stationary image classification, AURA shows substantial improvements in adaptation speed, accuracy, and computational efficiency over existing online learning and Bayesian filtering baselines, demonstrating that an adaptation-aware latent geometry is beneficial for effective Bayesian online learning in high-dimensional models.