Virtual Adversarial Training (VAT) is a regularization method for supervised and semi-supervised learning. It introduces a new measure of local smoothness of the conditional label distribution given input, called virtual adversarial loss. Unlike adversarial training, VAT defines the adversarial direction without label information, making it applicable to semi-supervised learning. VAT computes the gradient of virtual adversarial loss with minimal computational cost, using at most two forward-backward passes. Experiments show that VAT achieves state-of-the-art performance on semi-supervised tasks on SVHN and CIFAR-10. VAT is applicable to any parametric model where gradients can be computed with respect to input and parameters. It has few hyperparameters and is parameterization invariant. VAT is based on the idea of smoothing the model in the most anisotropic direction, quantified by the virtual adversarial direction. This direction is defined as the perturbation that most significantly alters the output distribution in terms of distributional divergence. VAT quantifies the local anisotropy of the model at each input point without using supervisory signals. VAT is a training method that maximizes the likelihood of the model while promoting local distributional smoothness (LDS) on each training input. It is applicable to semi-supervised learning and can be implemented with minimal computational cost. VAT is compared to random perturbation training (RPT) and shown to be more effective in reducing generalization error. VAT is also more effective in improving generalization performance when the true output distribution is isotropically smooth around input data points. Experiments on MNIST, CIFAR-10, and SVHN datasets show that VAT outperforms other methods in supervised and semi-supervised learning tasks. VAT achieves superior performance on these tasks, especially when combined with entropy minimization. VAT is also effective in improving the robustness of trained neural networks against virtual adversarial perturbations. The method is efficient, with only two hyperparameters, and is parameterization invariant. VAT is a promising approach for improving the generalization performance of models in both supervised and semi-supervised learning tasks.Virtual Adversarial Training (VAT) is a regularization method for supervised and semi-supervised learning. It introduces a new measure of local smoothness of the conditional label distribution given input, called virtual adversarial loss. Unlike adversarial training, VAT defines the adversarial direction without label information, making it applicable to semi-supervised learning. VAT computes the gradient of virtual adversarial loss with minimal computational cost, using at most two forward-backward passes. Experiments show that VAT achieves state-of-the-art performance on semi-supervised tasks on SVHN and CIFAR-10. VAT is applicable to any parametric model where gradients can be computed with respect to input and parameters. It has few hyperparameters and is parameterization invariant. VAT is based on the idea of smoothing the model in the most anisotropic direction, quantified by the virtual adversarial direction. This direction is defined as the perturbation that most significantly alters the output distribution in terms of distributional divergence. VAT quantifies the local anisotropy of the model at each input point without using supervisory signals. VAT is a training method that maximizes the likelihood of the model while promoting local distributional smoothness (LDS) on each training input. It is applicable to semi-supervised learning and can be implemented with minimal computational cost. VAT is compared to random perturbation training (RPT) and shown to be more effective in reducing generalization error. VAT is also more effective in improving generalization performance when the true output distribution is isotropically smooth around input data points. Experiments on MNIST, CIFAR-10, and SVHN datasets show that VAT outperforms other methods in supervised and semi-supervised learning tasks. VAT achieves superior performance on these tasks, especially when combined with entropy minimization. VAT is also effective in improving the robustness of trained neural networks against virtual adversarial perturbations. The method is efficient, with only two hyperparameters, and is parameterization invariant. VAT is a promising approach for improving the generalization performance of models in both supervised and semi-supervised learning tasks.