Using Variational approaches to perform computationally efficient 'divide and conquer' Monte Carlo inference on demographic models.

Simulator-based models are becoming ubiquitous in engineering and sciences because their mechanistic behaviour is easy to interpret and appropriately emulates natural phenomena. Epidemiology, population genetics, cosmology and computational neuroscience have developed complex simulations to make sense of our world, however, such simulator-based models make statistical analyses prohibitively intractable and expensive. Likelihood-free inference methods partially overcome the intractability but require vast amounts of simulations, which can often be expensive or time-consuming. On the other hand, modern machine learning methods have been shown to effectively scale to big data and achieve state-of-the-art results across a wide range of applications. My research will investigate how one can leverage scalable machine learning methods by combining neural networks with statistically-sound variational and likelihood-free inference to extract further insight from simulator-based models without trading-off computational resources. This venue of research is still in its infancy but has seen a lot of excitement and momentum in the Machine Learning community. My research would take Google DeepMind's latest research one step further and its success would boost scientific innovation and understanding of natural phenomena across all engineering and science disciplines, thus not only revolutionising the likelihood-free inference area of Machine Learning but, rather, it has the potential to have an impact on any discipline requiring a simulator-based model, and this is especially important nowadays because Epidemiologic simulations could help prevent and manage another pandemic.