Continual Learning from an Open-World Perspective

Abstract

Continual learning is an approach to deep learning that trains models on new data while preserving previously learned knowledge, much like a biological learner. One common source of change is time: people change, places change, facts change, and meanings change.

Adapting neural networks to change is costly and infrequent. With continual learning, models could stay up to date with knowledge cut-offs of days, minutes, or seconds rather than months or years. Unfortunately, deep neural networks suffer from catastrophic forgetting, where new information interferes with old knowledge.

We investigate continual learning from an open-world perspective to push continual learning strategies into more challenging, realistic settings that make fewer assumptions. The perspective arises from machine learning research directions that share the open-world problem. The open-world problem is that the training set never contains every possibility, so we must detect, characterise, and learn from the unknown. We argue that this open-world perspective is critical to robust, continual learning.

Our efforts culminate in four contributions. First, inspired by anomaly detection, we adapt task-incremental parameter-isolation strategies to class-incremental continual learning, thereby removing the assumption of task labels. Second, we investigate claims of resistance to catastrophic forgetting in the Kolmogorov-Arnold Networks architecture, finding that it still forgets when assumed simplistic relations between features are replaced with realistic datasets. Third, we propose a Bayesian adapter strategy to jointly address confidence calibration and continual learning, challenging the implicit assumption that these two challenges require separate mechanisms. Finally, building on drift detection, we move beyond the task boundary assumptions of many continual learners.

The promise of continual learning is up-to-date awareness, robustness, and improved efficiency for neural networks operating in changing environments. This thesis enables more efficient and robust continual learning in realistic settings by reducing the underlying assumptions typically made by continual learners.