Gen.Jl

Gen is a general-purpose probabilistic programming system that automates the implementation details of probabilistic inference algorithms. It provides building blocks for writing efficient inference algorithms tailored to specific models, automating tricky math and low-level implementation details. Gen supports hybrid algorithms that combine neural networks, variational inference, sequential Monte Carlo, and Markov chain Monte Carlo. Users can write custom inference algorithms without extending the compiler, using a flexible API that includes automatic differentiation and many other operations needed for model-based inference. Gen also supports efficient inference in models with stochastic structure through custom reversible jump and involutive MCMC algorithms. Gen is designed for researchers and engineers who need to build generative models and perform probabilistic inference. It is particularly suited for those working on Bayesian statistics, computational cognitive science, and any domain where uncertainty quantification is critical. The system allows users to write generative models, inference models, variational families, and proposal distributions using ordinary Julia code, with the option to migrate parts to specialized modeling languages for faster performance. What sets Gen apart is its programmable inference approach: instead of a black-box inference engine, Gen provides a flexible API that lets users implement an open-ended set of inference and learning algorithms. This includes operations beyond automatic differentiation, enabling custom hybrid algorithms that leverage both neural network speed and model-based accuracy. Gen's unique support for dynamic computation graphs and custom MCMC kernels makes it powerful for models with stochastic structure. Gen is free and open-source, created by the MIT Probabilistic Computing Project. It is implemented in Julia, with ongoing ports to other languages. The main author is Marco Cusumano-Towner, and the