Research and Academic Code
2022Multi-Task Dynamical Systems (pdf)
Alex Bird, Christopher K.I. Williams, Christopher Hawthorne. Journal of Machine Learning Research 23(230) 1-52 (2022).
Video examples of the Mocap work can be seen here
2021
PhD Thesis - Multi-Task Dynamical Systems (pdf)
Edinburgh University, 2021
Adaptive Mixture Importance Sampling (GitHub)
- (defined in the above thesis)
- This is a sequential Monte-Carlo algorithm for sampling from a distribution over time. It is specifically useful for generating samples from a posterior distribution as more information is gleaned as time evolves. It is surprisingly hard to generate enough diversity using existing sequential algorithms.
Video Generation of a Double Pendulum from a Multi-Task Dynamical model (GitHub)
- Using a GRU (RNN-type) model.
- While this is a novel implementation (from scratch), producing videos of double pendulums is not novel. What is interesting here is using the latent task variable to model the different trajectories implied by the chaotic system. (See e.g. the three video panels in the README.)
- More examples can be found here.
2019
Multi-Task Time Series Analysis applied to Drug Response Modelling (pdf)
Alex Bird, Christopher K. I. Williams, Christopher Hawthorne. In 22nd International Conference on Artificial Intelligence and Statistics (AISTATS) 2019. PMLR Vol 89.