Learning useful representations without supervision remains a key challenge in machine learning. In this paper, we propose a simple yet powerful generative model that learns such discrete representations. Our model, the Vector Quantised-Variational AutoEncoder (VQ-VAE), differs from VAEs in two key ways: the encoder network outputs discrete, rather than continuous, codes; and the prior is learnt rather than static. In order to learn a discrete latent representation, we incorporate ideas from vector quantisation (VQ). Using the VQ method allows the model to circumvent issues of ``posterior collapse'' -— where the latents are ignored when they are paired with a powerful autoregressive decoder -— typically observed in the VAE framework. Pairing these representations with an autoregressive prior, the model can generate high quality images, videos, and speech as well as doing high quality speaker conversion and unsupervised learning of phonemes, providing further evidence of the utility of the learnt representations.
A latent-variable model is introduced for text matching, inferring sentence representations by jointly optimizing generative and discriminative objectives. To alleviate typical optimization challenges in latent-variable models for text, we employ deconvolutional networks as the sequence decoder (generator), providing learned latent codes with more semantic information and better generalization. Our model, trained in an unsupervised manner, yields stronger empirical predictive performance than a decoder based on Long Short-Term Memory (LSTM), with less parameters and considerably faster training. Further, we apply it to text sequence-matching problems. The proposed model significantly outperforms several strong sentence-encoding baselines, especially in the semi-supervised setting.
In this paper we study how to learn stochastic, multimodal transition dynamics in reinforcement learning (RL) tasks. We focus on evaluating transition function estimation, while we defer planning over this model to future work. Stochasticity is a fundamental property of many task environments. However, discriminative function approximators have difficulty estimating multimodal stochasticity. In contrast, deep generative models do capture complex high-dimensional outcome distributions. First we discuss why, amongst such models, conditional variational inference (VI) is theoretically most appealing for model-based RL. Subsequently, we compare different VI models on their ability to learn complex stochasticity on simulated functions, as well as on a typical RL gridworld with multimodal dynamics. Results show VI successfully predicts multimodal outcomes, but also robustly ignores these for deterministic parts of the transition dynamics. In summary, we show a robust method to learn multimodal transitions using function approximation, which is a key preliminary for model-based RL in stochastic domains.
In practice, conditional variational autoencoders (CV AEs) perform conditioning by combining two sources of information which are computed completely independently; CV AEs first compute the condition, then sample the latent variable, and finally concatenate these two sources of information. However, these two processes should be tied together such that the model samples a latent variable given the conditioning signal. In this paper, we directly address this by conditioning the sampling of the latent variable on the CV AE condition, thus encouraging it to carry relevant information. We study this specifically for tasks that leverage with strong conditioning signals and where the generative models have highly expressive decoders able to generate a sample based on the information contained in the condition solely. In particular, we experiments with the two challenging tasks of diverse human motion generation and diverse image captioning, for which our results suggest that unifying latent variable sampling and conditioning not only yields samples of higher quality, but also helps the model to avoid the posterior collapse, a known problem of V AEs with expressive decoders.
This paper introduces a new encoder-decoder architecture that is trained to reconstruct images by disentangling the salient information of the image and the values of attributes directly in the latent space. As a result, after training, our model can generate different realistic versions of an input image by varying the attribute values. By using continuous attribute values, we can choose how much a specific attribute is perceivable in the generated image. This property could allow for applications where users can modify an image using sliding knobs, like faders on a mixing console, to change the facial expression of a portrait, or to update the color of some objects. Compared to the state-of-the-art which mostly relies on training adversarial networks in pixel space by altering attribute values at train time, our approach results in much simpler training schemes and nicely scales to multiple attributes. We present evidence that our model can significantly change the perceived value of the attributes while preserving the naturalness of images.