Few-shot density estimation lies at the core of current meta-learning (or ‘learning to learn’) research and is crucial for intelligent systems to be able to adapt quickly to unseen tasks. In this talk we will introduce generative query networks (GQN - published in Science this year), a generative model for few-shot scene understanding that learns to capture the main features of synthetic 3D scenes. In the second half of the talk we will cover neural processes (NPs), a generalisation of the GQN training regime to a wider range of tasks like regression and classification. NPs are inspired by the flexibility of stochastic processes such as Gaussian processes, but are structured as neural networks and trained via gradient descent. We show how NPs make accurate predictions after observing only a handful of training data points, yet scale to complex functions and large data sets.