Detect out-of-distribution data in lifelong learning without memory
Out-of-Distribution Detection in Continual Learning
Machine LearningComputer Vision and Pattern Recognition
Summary
Detecting unknown data while teaching computers new tasks one after another is very hard because computers often forget old tasks and struggle to recognize new, unfamiliar input. The authors propose a new method that avoids storing old data, which is usually needed. Their approach uses a special kind of network that naturally remembers past tasks better and adjusts to each task to tell known data apart from unknown data. They also introduce a new way to score how likely data is unknown without looking back at old data. Together, these ideas improve the computer's ability to spot unfamiliar data while still performing well on known tasks.
What this means in practice
- •For machine learning engineers: Improve systems that learn multiple tasks over time to recognize new unknown inputs without storing past data.
- •For autonomous vehicle developers: Enhance vehicle systems to detect unknown environmental conditions encountered after software updates without large memory needs.
Authors
Nimeshika Udayangani, Flora D. Salim, Christopher Leckie, Sarah Erfani
Abstract
Out-of-distribution detection in continual learning (OOD-CL) is highly challenging, as it requires (i) learning a sequence of tasks without catastrophic forgetting, and (ii) identifying unknown samples (OOD) without access to known (in-distribution (ID)) data from past tasks, where standard confidence-based OOD scores tend to degrade. Existing methods largely rely on a rehearsal buffer of past data or assume closed-world settings, limiting their scalability and reliability in real-world scenarios. In this paper, we propose a novel OOD detection framework for rehearsal-free continual learning, inspired by Kolmogorov-Arnold Networks (KAN), which mitigate catastrophic forgetting via their inherent local neuroplasticity. To better leverage KAN in OOD-CL, we introduce a task-adaptive classifier architecture, TA-KAN, which mitigates recency bias and enhances ID--OOD separability through task-specific control of activation locality. Furthermore, we propose a geometry-guided OOD score that complements TA-KAN classifier confidence without relying on past ID data. Our method significantly improves OOD detection while maintaining strong ID classification performance, achieving state-of-the-art results in OOD-CL.