This talk will introduce fundamentals of probabilistic programming, including new probabilistic programming-based AI moonshots being pursued by MIT in collaboration with Google and Meta. Capabilities will be grounded in new SOTA results in indoor 3D scene perception, large-scale medical database linkage, automated modeling of clinical trials, and automated econometric forecasting.
It will also briefly review emerging economic opportunities and security risks arising from the surprising robustness, data efficiency, and compute efficiency of probabilistic programs, as compared to current neural networks on their own.

This talk will review fundamental concepts and theoretical results relevant for four topics: computability of probabilistic inference; the efficiency of random variate generation; prospects for decentralized, planetary-scale probabilistic computation; and prospects for using probabilistic programs to learn to generate synthetic data.

As computing power increases, algorithms are increasingly used in decision making. However, many existing algorithms significantly change their outputs when inputs change or noise is added. As a result, the decisions are also subject to frequent changes, which can result in large costs and loss of trust. To cope with this situation, the speakers proposed a concept called "average sensitivity of algorithms" and have proposed algorithms with small average sensitivity for various problems such as graph-related problems, clustering, and decision tree learning. In this talk, I will introduce recent progress on average sensitivity of algorithms.

This talk will introduce a close connection between dually flat manifolds, which is known as the canonical structure in information geometry, and incidence algebras in order theory, and present its applications to machine learning. This approach allows us to flexibly design log-linear models on a directed acyclic graph, which include a number of machine learning problems such as learning of Boltzmann machines, matrix/tensor low-rank approximation, and blind source separation.

Control flow is a crucial notion of running programs, determining the order in which instructions are executed. Because manipulating control flow in a well-behaved manner makes programs concise, intuitive, and structured, programming languages often provide constructs to operate control flow, such as exception handling. This talk will introduce algebraic effect handlers, a recently emerging construct for abstraction of control flow manipulation. Algebraic effect handlers generalize exception handling to allow the resumption of computation and have several applications including state, backtracking, generators, and probabilistic programming.