Emmy

Probabilistic programming language embedded in Scala. It's focus is on scalability, with inference algorithms like variational bayes and the No-U-Turn sampler.

Features

Emmy provides a language for (bayesian) inference and learning on the JVM.
When constructing a model of the data, i.e. a generative process based on some parameters, an abstract syntax tree (AST) of Expressions is constructed.

Flexible data types

Data container types include various scala collections, to make their use as natural as possible within a scala machine learning pipeline. Values can be integer- (Int) or real-valued (Double). Distributions can be continuous (Normal, Gamma) or discrete (Categorical).

Automatic Differentiation

When optimizing approximations to the posterior distribution, automatic differentiation is used. This takes a large, error-prone, burden off the process of implementing a new model.

Compiling to a function

Since values and derivatives are calculated many times during Monte Carlo evaluations, these are compiled to a function. Memoization reduces multiple uses of the same value to a single evaluation.

Dynamic Graph Construction

To allow for maximum flexibility for inference, graphs can be dynamic - they are reconstructed for each mini-batch of data.

Documentation

For documentation and examples, see http://vanlankvelt.com/emmy.