PRINCIPLED LEARNING

Abstract

The thesis explores leveraging principles from human learning to enhance machine learning models. It delves into four key principles: continual learning, recency, similarity-based learning, and predictive coding. The first part introduces a continual learning method, in which an Agent learns continuously over time, accommodating delayed feedback. This strategy enables flexibility in learning and improves performance in tasks like image classification and captioning. The second part focuses on using task similarity to enhance multi-task learning. We experiment with how virtual tasks are helpful in improving the performance of real tasks. The third part introduces a method for anomaly detection in multi-variate time series data. By leveraging predictive coding, it effectively models temporal dependencies and correlations, outperforming existing anomaly detection methods. Lastly, continual learning is demonstrated beneficial in medical imaging, aiding in data security by facilitating model sharing across multiple sources with confidentiality constraints.