Adversarially Learned Inference (ALI) is a model that jointly learns a generation network and an inference network using an adversarial process. The generation network maps samples from stochastic latent variables to the data space, while the inference network maps training examples in data space to the space of latent variables. An adversarial game is played between these two networks and a discriminative network that is trained to distinguish between joint latent/data-space samples from the generative network and joint samples from the inference network. The model is shown to learn mutually coherent inference and generation networks through inspections of model samples and reconstructions, and to achieve performance competitive with other recent approaches on the semi-supervised SVHN task. ALI is similar to Generative Adversarial Networks (GANs), but differs in that the generator has two components: the encoder, which maps data samples to z-space, and the decoder, which maps samples from the prior to the input space. The discriminator is trained to distinguish between joint pairs of samples from the encoder and decoder. The generator is trained to fool the discriminator, i.e., to generate x,z pairs that are indistinguishable one from another. ALI is trained using an adversarial game where the generator minimizes the Jensen-Shannon divergence between the encoder and decoder joint distributions. The model is shown to achieve competitive results on semi-supervised SVHN classification. ALI is also shown to produce reconstructions that often faithfully represent more abstract features of the input images, while making mistakes in capturing exact object placement, color, style and (in extreme cases) object identity. These reconstructions suggest that the ALI latent variable representations are potentially more invariant to these less interesting factors of variation in the input and do not devote model capacity to capturing these factors. ALI is also shown to be well suited to semi-supervised tasks, as it tends to focus on semantic information at the expense of low-level details. The model is trained on four different datasets, namely CIFAR10, SVHN, CelebA, and a center-cropped version of the ImageNet dataset. The model is shown to achieve competitive performance on the semi-supervised SVHN classification task. The results are presented in Table 1. The SDGM’s performance is achieved via a carefully designed two-layer architecture that explicitly takes label information into account in learning the representation. We expect that ALI would also gain by taking account of label information in learning the representation.Adversarially Learned Inference (ALI) is a model that jointly learns a generation network and an inference network using an adversarial process. The generation network maps samples from stochastic latent variables to the data space, while the inference network maps training examples in data space to the space of latent variables. An adversarial game is played between these two networks and a discriminative network that is trained to distinguish between joint latent/data-space samples from the generative network and joint samples from the inference network. The model is shown to learn mutually coherent inference and generation networks through inspections of model samples and reconstructions, and to achieve performance competitive with other recent approaches on the semi-supervised SVHN task. ALI is similar to Generative Adversarial Networks (GANs), but differs in that the generator has two components: the encoder, which maps data samples to z-space, and the decoder, which maps samples from the prior to the input space. The discriminator is trained to distinguish between joint pairs of samples from the encoder and decoder. The generator is trained to fool the discriminator, i.e., to generate x,z pairs that are indistinguishable one from another. ALI is trained using an adversarial game where the generator minimizes the Jensen-Shannon divergence between the encoder and decoder joint distributions. The model is shown to achieve competitive results on semi-supervised SVHN classification. ALI is also shown to produce reconstructions that often faithfully represent more abstract features of the input images, while making mistakes in capturing exact object placement, color, style and (in extreme cases) object identity. These reconstructions suggest that the ALI latent variable representations are potentially more invariant to these less interesting factors of variation in the input and do not devote model capacity to capturing these factors. ALI is also shown to be well suited to semi-supervised tasks, as it tends to focus on semantic information at the expense of low-level details. The model is trained on four different datasets, namely CIFAR10, SVHN, CelebA, and a center-cropped version of the ImageNet dataset. The model is shown to achieve competitive performance on the semi-supervised SVHN classification task. The results are presented in Table 1. The SDGM’s performance is achieved via a carefully designed two-layer architecture that explicitly takes label information into account in learning the representation. We expect that ALI would also gain by taking account of label information in learning the representation.