Joshua Goncalves (HDR @CSE)

A Hierarchical Approach to Accurate yet Explainable Artificial Intelligence

Abstract

Recent advancements in artificial intelligence have enabled plausible and effective solutions to a variety of applications including self-driving cars and real-time medical diagnosis. Many different algorithms and techniques have been utilised to achieve such results, driven by both commercial and academic pressure. The most prominent amongst them are arguably those classified as deep learning which use a hierarchical structure that progressively abstracts constituent elements until a more usable form is reached. Whilst their predictive power is quite significant, they generally operate in a somewhat opaque fashion such that their mainly quadratic internals cannot be interpreted in any meaningful way, leaving their conclusions arguably unsubstantiated

Symbolic approaches in comparison are capable of this necessary explanatory prowess. They work by representing various components through logic whose inherent form and shared relationships are rather explicit and therefore easily interpretable. As a consequence of this thorough detailing however, modelling large, complex phenomena can be unrealistic or considerably more difficult

We aim to design, build and validate a novel machine learning system that will comprise both hierarchical and symbolic structures. They will work in cooperation to model some visual phenomena in an interpretable fashion, developing jointly to achieve the same goal.

Bio

Joshua Goncalves is currently a PhD student in the School of Computer Science and En- gineering at the University of New South Wales. His research interests include machine learning and its application in computer vision, revolving mainly around explainability. Prior to his postgraduate degree, Joshua worked as a software and robotics engineer for the Aus- tralian Institute of Robotic Orthopaedics and was a visiting research associate at the Institute for Laser Technologies in Medicine and Measuring Technology (ILM, Germany).