Zhe Hou, Hadrien Bride & Jim Song Dong - Griffith University 

Machine learning for dependable decision-making

Abstract

Machine learning (ML) has been very successful in prediction, classification, regression, anomaly detection and other forms of data analytics.  ML is becoming an integrated part of automated decision-making for critical systems.  However, most existing ML techniques are used as black-boxes and they do not provide a high level of trustworthiness. In fact, there have been numerous cases where ML-based applications failed and caused tremendous damage.  To address the security and safety concerns of critical systems, advances in trust-related aspects of machine learning are important.  Explainable artificial inteligence (XAI) is one way to improve trustworthiness and it has attracted much attention recently.  Machine learning approaches that are capable of explaining the rationale behind the predictions are more relatable and transparent.  Another way to improve trustworthiness is to develop ML techniques that can produce auditable predictive models.  The verification of these models provides formal guarantee that the models are correct, safe and secure with respect to user's requirements.

In this tutorial we will cover recent developments in the domain of transparent and auditable machine learning techniques.  We will introduce our latest research combining advanced machine learning, high performance computing and automated reasoning techniques.  We will also present the fruit of our work:  Silas -- a state of the art ML toolkit for dependable machine learning.