Grant: Discovery Grants Program - Individual

Human beings are capable of continually acquiring, correcting, and transferring knowledge and skills throughout our lifetimes. We do so by multi-tasking and learning from the past while constantly improving our abilities based on newly acquired knowledge and life experiences. The goal of lifelong machine learning systems is to mimic this behaviour. To do so, algorithms are required to (a) maintain accurate models about multiple tasks in parallel, (b) fine-tune existing models when more information about known tasks arrives, (c) create new models to handle new tasks, (d) reuse models when tasks reappear, and (e) construct accurate models for highly skewed distributions - a frequent occurrence in real-world domains where essential concepts are represented by very few examples. As such, lifelong learners should effectively identify and transfer relevant knowledge between models and across (related) domains. In online lifelong learning, also known as incremental or streaming learning, models incrementally learn from new instances without requiring access to past data, facilitating continuous learning in near real-time environments. Online lifelong learning algorithms offer advances in various domains. In industrial settings, such as online predictive maintenance, models can continuously detect equipment failures, adapt to changing environmental conditions, and improve the efficiency of maintenance scheduling. In healthcare, systems utilizing individual patient health data can improve predictions and alert systems, enabling personalized treatment plans, disease prediction, and improved patient outcomes. Online lifelong learning approaches are invaluable in the financial sector to aid in identifying new types of fraudulent activities in online fraud detection models, which would minimize false positives while ensuring that the system remains effective as new fraud methods emerge. In adaptive intrusion detection systems developed for cybersecurity, online lifelong learning can be used to continuously learn and adapt to new evasion techniques employed by attackers. Designing accurate, timely online lifelong learning algorithms is a fundamental topic of broad interest and is the primary objective of this research program. Our work will focus on algorithms that enable online knowledge transfer while handling reoccurring tasks and emerging concepts. We will create solutions to mitigate the effects of data skew (or imbalance). Throughout this program, we will design and deploy accurate yet transparent algorithms that adhere to Responsible AI principles. This adherence ensures we minimize the risk of exacerbating societal biases and inequalities when deploying our novel solutions.