Probabilistic AI
Methods grounded in probability theory and statistical inference. Bayesian reasoning, graphical models, and uncertainty quantification.
Sub-topics
Directed acyclic graphs representing conditional dependencies. Judea Pearl formalized them in the 1980s. Enable principled reasoning under uncertainty.
Statistical models for sequential data with hidden states. Dominated speech recognition from the 1980s through the 2010s. Also used in bioinformatics and NLP.
Non-parametric Bayesian models defining distributions over functions. Provide uncertainty estimates naturally. Used in Bayesian optimization and surrogate modeling.
Family of algorithms for sampling from probability distributions. Metropolis-Hastings (1953/1970) and Gibbs sampling enable Bayesian inference in complex models.
Approximate Bayesian inference by optimization rather than sampling. Turns intractable posterior computation into a tractable optimization problem. Faster than MCMC.
Latent Dirichlet Allocation (Blei et al., 2003) discovers latent topics in document collections using Bayesian generative models. Influential in text mining and NLP.
Languages that combine programming with probabilistic modeling. Stan, Pyro, and NumPyro allow specifying generative models and performing automatic inference.
Judea Pearl's do-calculus and causal diagrams formalize reasoning about cause and effect beyond correlation. A bridge between statistics and interventional reasoning.