edward2 - A Probabilistic Programming Language for Researchers and Developers in AI and Deep Learning

Introduction to Edward2

Edward2 is a simple yet powerful probabilistic programming language developed to work seamlessly within the deep learning ecosystem. It allows users to write models as probabilistic programs and gives them the flexibility to manipulate a model's computation for both training and inference purposes. Edward2 is designed to be easy to use while providing strong functionalities, making it ideal for those interested in probabilistic models.

Structure of Edward2

Edward2 is well-structured, consisting of:

Library Code: The core codebase that powers Edward2. Explore the Edward2 Library.
Examples: A collection of examples demonstrating how to use Edward2 effectively. View Examples.
Experimental Projects: Contains ongoing research projects using Edward2. Check out Experimental Projects.

For those who have previously worked with the original Edward, transitioning to Edward2 is made easier with a dedicated upgrading guide.

Features and Utilities

1. Models as Probabilistic Programs

Random Variables: Edward2 provides the RandomVariables class to define the structure of probabilistic models. A random variable in Edward2 is linked to a probability distribution managed as a TensorFlow Distribution instance. This allows for operations like log_prob and sample.
Probabilistic Models: These are expressed through Python functions that create RandomVariables. Such models can represent a wide range of data and distribution mechanisms, orchestrating processes like Bayesian logistic regression.

2. Manipulating Model Computation

Tracing: With Edward2, tracing assists users in training and testing probabilistic models by manipulating their computations. This flexibility is crucial for tasks like making predictions with learned posterior means rather than relying on priors.
Program Transformations: These involve mapping one model representation to another, granting access to various model properties. Program transformations are particularly useful for implementing algorithms like Markov chain Monte Carlo, which require a model's log-joint probability function.

Backend Support and Installation

Edward2 provides flexibility by supporting multiple backends, including TensorFlow, JAX, and NumPy. Users can activate specific backends by modifying their installation commands to include the necessary dependencies.

To use Edward2, installation can be done via pip using either the latest development version or the stable version (though updates to stable versions are infrequent due to its part-time development nature).

pip install "edward2 @ git+https://github.com/google/edward2.git"  # Latest development version
pip install edward2[tensorflow]  # For TensorFlow backend and dependencies

Conclusion

Edward2 stands out as an intuitive and effective tool for creating and managing probabilistic models. Its extensibility through different backends, coupled with its comprehensive support for probabilistic programming, makes it a valuable resource for practitioners in AI and machine learning fields.

For further knowledge about Edward2's application in different domains or additional resources, users are encouraged to explore its references and citations for deep dives into its methodologies and impacts in neural information processing systems.