Sherjil Ozair

Describes a generative interactive environment trained from unlabelled Internet video that can be prompted to generate action‑controllable virtual worlds; positioned as a foundation world model and co‑authored by multiple DeepMind researchers including Sherjil Ozair.

Work presenting discrete autoencoder and vector‑quantized approaches for planning in stochastic and partially observable environments (applies to model‑based RL and planning).

Paper addressing limitations of mutual-information based representation learning and proposing a practical approximation using Lipschitz constraints from the GAN literature to improve learned representations.

Introduces a neural estimator for mutual information scalable by gradient descent; used widely in representation learning and related research.

Foundational paper introducing Generative Adversarial Networks (GANs), a framework for training generative models via a two‑player adversarial process. (Sherjil Ozair is a co‑author on later versions/editions and related GAN works and commentary).