Research map / Continual learning / Online and replay-based
Online Learning via Learned Latent Bayesian Tracking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AURA, a meta-learned low-dimensional latent state-space model with EKF tracking and a learned lifting map to parameters, enabling single-step online adaptation for classifiers and neural wireless receivers.
Problem
Bayesian filtering for online adaptation of deep models is computationally prohibitive in full parameter space, and existing methods rely on restrictive covariance approximations or manually chosen subspaces.
Approach
Perform EKF tracking in a learned low-dimensional latent space governed by a diagonal linear state-space model, with the full parameters recovered by a learned lifting network, and meta-learn the lifting map and dynamics end-to-end on offline non-stationary trajectories.
Key results
- On MNIST-C with 15 samples per step, AURA reaches 1.63% error vs 2.64% for the best batch-wise baseline ROID, and CIFAR-10-C 14.28% vs 16.00%.
- On CIFAR-100-C, ROID (26.50%) is best, AURA (26.66%) second, versus Online GD 33.22% and EKF-FC 32.43% at 15 samples.
- In wireless receiver experiments, AURA reports lowest BER across the SNR sweep for linear and nonlinear QuaDRiGa channels, outperforming full-parameter BONG variants.
Surprising findings
- A streaming-feasible latent Bayesian method outperforms batch-wise test-time adaptation methods on MNIST-C and CIFAR-10-C despite using single-sample updates.
- Increasing latent dimension beyond a point gives saturating or slightly worse performance due to over-parameterization or estimation noise.
Limitations
- Theoretical motivation relies on strong assumptions (strong quasi-convexity, Lipschitz compressibility) that may not hold in deep learning.
- On CIFAR-100-C, AURA is not best; ROID attains lower error, and ROID is not streaming-compatible.
- Bayesian posterior is only over latent state, not full model weights, so no full uncertainty quantification over network weights.
Experimental setup
Datasets: MNIST-C CIFAR-10-C CIFAR-100-C MNIST CIFAR-10 CIFAR-100 QuaDRiGa
Benchmarks: MNIST-C CIFAR-10-C CIFAR-100-C QuaDRiGa linear and nonlinear channel BER
Metrics: classification error bit error rate FLOPs per update
Baselines: Online GD EKF-FC BN-Test BN-EMA BN-Alpha ROID BONG
Builds on: extended Kalman filter BONG DeepSIC ROID
Closest papers by meaning
- Kalman Filter for Online Classification of Non-Stationary Data (ICLR 2024 · similarity 0.86)
- Stationary Latent Weight Inference for Unreliable Observations from Online Test-Time Adaptation (ICML 2024 · similarity 0.86)
- Online Bernstein-von Mises theorem (NeurIPS 2026 · similarity 0.85)
- Bayesian Power Steering: An Effective Approach for Domain Adaptation of Diffusion Models (ICML 2024 · similarity 0.85)
- Debiased Offline Representation Learning for Fast Online Adaptation in Non-stationary Dynamics (ICML 2024 · similarity 0.84)
- Bayesian Online Natural Gradient (BONG) (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.