Online Continual Learning on Hierarchical Label Expansion

Continual learning (CL) enables models to adapt to new tasks and environments without forgetting previously learned knowledge. While current CL setups have ignored the relationship between labels in the past task and the new task with or without small task overlaps, real-world scenarios often involve hierarchical relationships between old and new tasks, posing another challenge for traditional CL approaches. To address this challenge, we propose a novel multi-level hierarchical class incremental task configuration with an online learning constraint, called hierarchical label expansion (HLE). Our configuration allows a network to first learn coarse-grained classes, with data labels continually expanding to more fine-grained classes in various hierarchy depths. To tackle this new setup, we propose a rehearsal-based method that utilizes hierarchy-aware pseudo-labeling to incorporate hierarchical class information. Additionally, we propose a simple yet effective memory management and sampling strategy that selectively adopts samples of newly encountered classes. Our experiments demonstrate that our proposed method can effectively use hierarchy on our HLE setup to improve classification accuracy across all levels of hierarchies, regardless of depth and class imbalance ratio, outperforming prior state-of-the-art works by significant margins while also outperforming them on the conventional disjoint, blurry and i-Blurry CL setups.

1. Introduction

In real-world continual learning scenarios, new knowledge often augments existing understanding, typically following a hierarchical path from general to specific classes. This hierarchical structure is not an anomaly, but rather an inherent part of many disciplines. The schema theory [9]; [43] in cognitive psychology and the conceptual clustering theory [28] in machine learning both emphasize hierarchical organization of knowledge. The COBWEB algorithm [20], a prominent machine learning method, uses hierarchical clustering for grouping related instances into meaningful categories. Hierarchical organization is also observed in biology’s taxonomy theory [8], classifying organisms based on shared traits, and in chemistry [27], where elements are arranged hierarchically according to their atomic properties. However, despite the prevalence of hierarchical relationships in these areas, many previous continual learning works [3]; [5]; [6]; [31] do not fully incorporate these relationships. This may be an area that needs more attention, as hierarchical relationships could play a role in knowledge evolution in incremental learning.