edward

Why we included this project

Researchers and statisticians who want to add uncertainty to deep learning will find Edward useful. You express a model as a probabilistic program, then fit it with a range of inference algorithms rather than a single fixed training loop. Models are built from familiar pieces like directed graphical models, neural networks through TensorFlow and Keras, and Bayesian nonparametrics, and you can choose between variational methods, Monte Carlo samplers such as Hamiltonian Monte Carlo, or combined schemes like expectation-maximization. That flexibility is most valuable when you are prototyping a new idea and want to see how different inference approaches behave on the same model. Edward runs on TensorFlow's computational graph, so you keep automatic differentiation, GPU and CPU execution, and distributed training, but it is a research-oriented library rather than a managed service, which suits teams comfortable wiring things together themselves.