blei-lab/edward

Features

• Bayesian Neural Networks - Implements neural networks with prior distributions on weights and biases to treat parameters as latent variables.
• Deep Generative Model Construction - Enables building and training complex neural networks that model data distributions using probabilistic programming in TensorFlow.
• Reparameterization Trick - Utilizes the reparameterization trick to isolate randomness, allowing gradients to flow through stochastic nodes in the computational graph.
• Bayesian Inference - Provides a comprehensive system for estimating parameters and latent variables through MAP estimation, Gibbs sampling, and amortized inference.
• Bayesian Neural Networks - Implements neural networks with prior distributions on weights to treat parameters as latent variables.
• Computational Graph Frameworks - Represents probabilistic models as differentiable computational graphs within the TensorFlow framework for automatic gradient calculation.
• Deep Generative Modeling Frameworks - Provides a toolkit for constructing Bayesian neural networks and mixture models to represent complex data distributions.
• Markov Chain Monte Carlo Sampling - Implements Markov Chain Monte Carlo methods including Gibbs sampling to approximate complex posterior distributions.
• Gibbs Samplers - Executes Markov chain Monte Carlo by iterating over conditional draws to approximate complex posterior distributions.
• Maximum A Posteriori Estimators - Optimizes model parameters by maximizing the posterior distribution using training data and iterative loss updates.
• Posterior Predictive Checks - Analyzes model fit by comparing test statistics from generated data against real observed data.
• Posterior Inference - Serves as an engine for approximating posterior distributions using stochastic gradients, Gibbs sampling, and amortized variational inference.
• Amortized Inference Networks - Executes inference using amortized computation via inference networks to train and test deep generative models.
• Variational Parameter Mapping - Maps data points to local variational parameters via a neural network to enable fast inference on new data.
• Variational Inference Implementations - Computes posterior distributions by minimizing KL divergence against simpler distribution families using stochastic gradients.
• Variational Posterior Approximations - Optimizes variational distributions to minimize Kullback-Leibler divergence from intractable true posterior distributions.
• Probabilistic Model Inference - Estimates parameters and latent variables using variational inference, Monte Carlo methods, or generative adversarial networks.
• Probabilistic Modeling - Constructs Bayesian neural networks, mixture models, and Gaussian processes for supervised and unsupervised learning tasks.
• Probabilistic Models - Defines statistical relationships between latent variables and observed data using priors and likelihood functions.
• Stochastic Gradient Estimators - Computes unbiased gradients of the evidence lower bound using reparameterization or score function estimators.
• Variational Inference - Approximates intractable posterior distributions by optimizing simpler models to minimize divergence via stochastic gradients.
• Probabilistic Programming Languages - Functions as a probabilistic programming language built on TensorFlow for deep generative modeling and variational inference.
• Bayesian Probabilistic Programming Frameworks - Provides an integrated environment for defining statistical relationships between latent variables and observed data to create Bayesian networks.
• Maximum A Posteriori Estimators - Finds the most likely configuration of hidden patterns by optimizing joint density or minimizing cross entropy.
• Gaussian Processes - Models nonlinear relationships by defining distributions over functions to perform binary classification with uncertainty.
• Generative Model Evaluation - Validates the performance of generative models using posterior predictive checks and data fit visualizations.
• Generative Adversarial Network Training - Implements a bilevel optimization process to train generative adversarial networks for complex data distribution approximation.
• Batch Parameter Updates - Enables updating model parameters using small data samples to improve processing speed and scale inference.
• Mixture Model Estimation - Assigns data points to distinct components using mixing probabilities and individual distributions to perform clustering.
• Mixture Density Networks - Combines neural networks with mixture density models to parameterize membership probabilities for complex data.
• Model Evaluation Metrics - Calculates quantitative metrics such as classification error and mean absolute error to assess trained models.
• Model Fit Analysis - Evaluates model accuracy by calculating residuals between predicted and observed values to verify distribution fit.
• Model Fit Visualizations - Conducts graphical checks and visualizes data fit by drawing samples from an inferred model to assess performance.
• Model Performance Evaluators - Implements tools for quantifying the quality and reliability of probabilistic models through point-based evaluations.
• Model Prediction Evaluation - Calculates prediction error on held-out testing data using the posterior predictive distribution.
• Distribution Analysis - Generates density plots and samples from learned distributions to verify if the model captures true data patterns.
• Generative Model Criticism - Evaluates the fit and predictive performance of generative models to determine if they capture underlying data processes.
• Minibatch Training Utilities - Processes large datasets by training on small subsets to optimize memory and compute resources.
• ELBO Gradient Estimation - Calculates unbiased gradients of the evidence lower bound using score function estimators or reparameterization.
• KL Divergence Gradient Estimation - Calculates noisy gradient estimates of the KL divergence by drawing and weighting samples from a variational distribution.
• Constrained Variable Transformations - Provides automated transformations to map constrained continuous random variables to an unconstrained space for gradient-based inference.
• Probabilistic Modeling - Probabilistic modelling, inference, and criticism.

Star history