Differentiable Physical Models for Data Analysis in Biology
Modern machine learning methods, such as deep neural networks, have delivered impressive results in recent years but suffer from the limitation that very large data sets are needed to train the models, limiting their applicability in many domains. My project focuses on augmenting neural networks with physical simulations, allowing real-world knowledge to be built into neural network architectures, which would otherwise need to learn everything they know directly from data. By combining neural networks and physical simulations I combine the best of two worlds: neural networks can model unknowns of a system, while physical simulation can provide real-world “expert” input. The hope is that this will allow a high degree of generalisability to be achieved from small datasets. I will apply the new methods to study physical and biological phenomena such as bacterial growth, microorganism motility, protein aggregation, and organ development.