Bayesian Online Learning
Bayesian Online Learning (Sequential Posterior Update) · Also known as: online Bayesian inference, sequential Bayesian learning, recursive Bayesian estimation, BOL
Bayesian online learning applies Bayesian inference sequentially: each time a new observation arrives, the current posterior over model parameters becomes the prior for the next update. The result is a principled probabilistic framework that maintains calibrated uncertainty estimates throughout, making it well-suited for streaming and non-stationary data settings.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Bayesian online learning when data arrive as a stream and re-training from scratch is infeasible or too slow; when calibrated uncertainty estimates are required alongside predictions (e.g., safety-critical or financial applications); or when the data-generating process may drift gradually over time and the model must adapt continuously. It is particularly appropriate for classification and regression with conjugate or approximately conjugate likelihoods. Avoid it when data are fully available up front and exact Bayesian inference is tractable — standard batch Bayesian methods will be more accurate. Also avoid it when the likelihood is highly non-conjugate and computation budget is very tight, as approximation errors can accumulate over many updates.
Strengths & limitations
- Maintains calibrated predictive uncertainty at every step, not just a point forecast.
- Does not require storing historical data — the posterior summary is sufficient for future updates.
- Adapts naturally to concept drift by continuously revising parameter beliefs.
- Principled framework: predictions follow from probability theory, easing interpretation for reviewers.
- Conjugate models admit exact, closed-form updates with negligible computational cost per step.
- Regularisation emerges automatically from the prior, reducing overfitting risk.
- Choosing a well-specified prior requires domain expertise; a poor prior can bias results for many steps before enough data correct it.
- Non-conjugate likelihoods force approximations (variational Bayes, EP, Laplace) that introduce their own errors.
- Approximate methods can accumulate errors over long data streams, causing the posterior to drift from the true posterior.
- Computationally more demanding per update than simple stochastic gradient methods for large parameter spaces.
Frequently asked
Is Bayesian online learning the same as the Kalman filter?
The Kalman filter is a special case of Bayesian online learning where the likelihood and transition model are linear-Gaussian, enabling exact closed-form posterior updates. Bayesian online learning is the broader framework that encompasses non-linear and non-Gaussian settings through approximations.
What happens if I use the wrong prior?
A misspecified prior biases the posterior for the first many observations. With enough data, the likelihood dominates and the posterior concentrates near the true parameter regardless of the prior. However, in high-dimensional settings or under concept drift this correction may be slow or incomplete.
How do I handle non-conjugate likelihoods?
Common strategies include Laplace approximation (Gaussian approximation around the mode), online variational Bayes (optimise a tractable lower bound), expectation propagation (match moments of the posterior), and particle filters (Monte Carlo approximation). Each trades accuracy for computational cost differently.
Can Bayesian online learning handle concept drift?
Standard formulations assume a stationary parameter, so sharp concept drift is poorly handled. Forgetting mechanisms — such as discounting the prior with a leakage factor or using a dynamic Bayesian model with a random-walk transition — extend the framework to non-stationary environments.
How is this different from plain stochastic gradient descent?
SGD finds a single point estimate of the parameters by following noisy gradient steps; it produces no uncertainty estimate. Bayesian online learning maintains a full distribution over parameters, enabling calibrated predictive intervals at the cost of higher computational complexity per update.
Sources
- Opper, M. (1998). A Bayesian approach to on-line learning. In D. Saad (Ed.), On-Line Learning in Neural Networks (pp. 363–378). Cambridge University Press. link ↗
- Sato, M. (2001). Online model selection based on the variational Bayes. Neural Computation, 13(7), 1649–1681. DOI: 10.1162/089976601750265045 ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Online Learning (Sequential Posterior Update). ScholarGate. https://scholargate.app/en/machine-learning/bayesian-online-learning
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian Gaussian ProcessMachine learning↔ compare
- Bayesian Logistic RegressionBayesian↔ compare
- Gaussian ProcessMachine learning↔ compare
- Online LearningMachine learning↔ compare
- Semi-supervised LearningMachine learning↔ compare
- Variational InferenceBayesian↔ compare