Will Harvey

I am a final-year PhD student at the University of British Columbia, supervised by Frank Wood. Previously I've interned at Google DeepMind and studied for an MEng in Engineering Science at the University of Oxford.

My research focuses on generative modeling, especially with diffusion models. I investigate applications both in the video domain and for problems with more explicit structural information.

Email  /  Google Scholar  /  CV

We generalise diffusion models to be applicable to data with varying dimensionality, like molecules with varying numbers of atoms. We use jump diffusion to simultaneously sample the dimensionality along with the state.

A framework for incorporating known problem structure in diffusion model design. Doing so is necessary to scale to large problem sizes on problems including matrix factorisation and Sudoku-solving.

Conditional image diffusion models typically outperform unconditional diffusion models, and we show that the gap between them grows as more information is provided to condition on. We then present a simple method to close this gap, yielding improved unconditional and class-conditional image generation.

We present a diffusion model for video. Given a few frames of context it can generate videos over an hour long.

We present conditional VAEs which can be quickly trained by leveraging existing unconditional VAEs. The resulting models provide a more faithful representation of uncertainty than GAN-based approaches with similar training times.

We demonstrate how existing software tools can be used to automate parts of infectious disease-control policy-making via performing inference in existing epidemiological dynamics models.

Attention for Inference Compilation
William Harvey*, *, , ,
SIMULTECH 2022

We present a transformer-based architecture for improved amortized inference in probabilistic programs with complex and stochastic control flow.

Bayesian experimental design can be used to find near-optimal attention locations for a hard attention mechanism. These can be used to speed up the later training of hard attention mechanisms.

We improve image quality by training a GAN generator in a way that accounts for a sub-optimal discriminator.

We use knowledge about the structure of a generative model to automatically select a good normalizing flow architecture.

We present an interface for gradient-based training of pipelines of machine learning primitives. This allows joint training of machine learning modules written in different languages, making it useful for automated machine learning (AutoML).