Auto-Encoding Variational Bayes (AEVB) is a method for efficient inference and learning in directed probabilistic models with continuous latent variables and intractable posterior distributions. The paper introduces a stochastic variational inference and learning algorithm that can scale to large datasets and works even in the intractable case under mild differentiability conditions. The key contributions are: (1) a reparameterization of the variational lower bound that allows for a differentiable, unbiased estimator of the lower bound, which can be optimized using standard stochastic gradient methods; and (2) an efficient algorithm for posterior inference and learning, AEVB, which uses this estimator to fit an approximate inference model (recognition model) to the intractable posterior. The recognition model allows for efficient approximate posterior inference using simple ancestral sampling, which in turn enables efficient learning of the model parameters without the need for expensive iterative inference schemes. Theoretical advantages are reflected in experimental results. The method is applicable to a wide range of models, including those with i.i.d. datasets and continuous latent variables per datapoint. The paper also discusses related work, including the wake-sleep algorithm and stochastic variational inference, and presents experiments comparing AEVB to these methods on image datasets. The results show that AEVB converges faster and achieves better performance than other methods. The paper concludes that AEVB provides a novel estimator of the variational lower bound, SGVB, for efficient approximate inference with continuous latent variables, and introduces an efficient algorithm for inference and learning, AEVB, that learns an approximate inference model using the SGVB estimator. The theoretical advantages are reflected in experimental results.Auto-Encoding Variational Bayes (AEVB) is a method for efficient inference and learning in directed probabilistic models with continuous latent variables and intractable posterior distributions. The paper introduces a stochastic variational inference and learning algorithm that can scale to large datasets and works even in the intractable case under mild differentiability conditions. The key contributions are: (1) a reparameterization of the variational lower bound that allows for a differentiable, unbiased estimator of the lower bound, which can be optimized using standard stochastic gradient methods; and (2) an efficient algorithm for posterior inference and learning, AEVB, which uses this estimator to fit an approximate inference model (recognition model) to the intractable posterior. The recognition model allows for efficient approximate posterior inference using simple ancestral sampling, which in turn enables efficient learning of the model parameters without the need for expensive iterative inference schemes. Theoretical advantages are reflected in experimental results. The method is applicable to a wide range of models, including those with i.i.d. datasets and continuous latent variables per datapoint. The paper also discusses related work, including the wake-sleep algorithm and stochastic variational inference, and presents experiments comparing AEVB to these methods on image datasets. The results show that AEVB converges faster and achieves better performance than other methods. The paper concludes that AEVB provides a novel estimator of the variational lower bound, SGVB, for efficient approximate inference with continuous latent variables, and introduces an efficient algorithm for inference and learning, AEVB, that learns an approximate inference model using the SGVB estimator. The theoretical advantages are reflected in experimental results.