Universal Learnt Representations in Artificial Intelligence

Despite recent successes of artificial intelligence (AI) models for vision and language tasks, we do not have a good sense of how they fully work. Their non-linear nature makes them hard to interpret and model mathematically. A good example is human vision, where biological experiments in other mammals have shown that special functions called receptive fields are present in the visual cortex of the brain. Neural networks end up developing these same functions in an independent (data driven) manner despite not providing them any knowledge of this function beforehand. Another example is how neural networks seem to learn underlying representations of their training data that are equivalent to curved surfaces called manifolds, which is also a key mathematical concept in our understanding of gravity with General Relativity. This manifold representation has strong empirical evidence, but its exact nature is still unknown and is called the manifold hypothesis. In this project, we will explore such universal representations in AI with the candidate having the flexibility in choosing these application areas. The project will eventually develop theories that could help us understand and explain this universality in neural networks and AI in general.

The successful candidate will join ARC Future Fellow Dr Shekhar S. Chandra’s team of AI researchers at the University of Queensland. Dr Chandra is an experienced supervisor and leading researcher in imaging and AI, covering diverse areas such as vision, medical imaging, multi-modal & metric learning, chaos theory & fractals, signal processing and number theory. He also has strong research and industry ties with partners such as CSIRO Australia and Siemens Healthineers (Erlangen, Germany and Brisbane, Australia). UQ’s Research Computing Centre will provide state-of-the-art high-performance computing that will be essential in building the necessary AI models for this project.