[Paper Review] Continual Learning of a Mixed Sequence of Similar and Dissimilar Tasks

Zixuan Ke, Bing Liu+1 more|arXiv (Cornell University)|Dec 18, 2021

Domain Adaptation and Few-Shot LearningComputer Science41 references53 citations

TL;DR

CAT learns mixed sequences of similar and dissimilar tasks by protecting dissimilar-task knowledge, selectively transferring knowledge from similar tasks, and automatically detecting task similarity, enabling forgetting avoidance and forward/backward knowledge transfer within a single network.

ABSTRACT

Existing research on continual learning of a sequence of tasks focused on dealing with catastrophic forgetting, where the tasks are assumed to be dissimilar and have little shared knowledge. Some work has also been done to transfer previously learned knowledge to the new task when the tasks are similar and have shared knowledge. To the best of our knowledge, no technique has been proposed to learn a sequence of mixed similar and dissimilar tasks that can deal with forgetting and also transfer knowledge forward and backward. This paper proposes such a technique to learn both types of tasks in the same network. For dissimilar tasks, the algorithm focuses on dealing with forgetting, and for similar tasks, the algorithm focuses on selectively transferring the knowledge learned from some similar previous tasks to improve the new task learning. Additionally, the algorithm automatically detects whether a new task is similar to any previous tasks. Empirical evaluation using sequences of mixed tasks demonstrates the effectiveness of the proposed model.

Motivation & Objective

• Motivate continual learning to handle both forgetting and knowledge transfer in a mixed sequence of tasks.
• Propose a TCL model (CAT) that uses a shared knowledge base and task-specific masks to prevent forgetting.
• Enable selective forward knowledge transfer from similar tasks via a knowledge transfer attention mechanism.
• Allow backward knowledge transfer to improve similar previous tasks while learning a new task.
• Automatically detect task similarity to determine which previous tasks can transfer knowledge to the current task.

Proposed method

• Introduce a knowledge base (KB) shared by all tasks to store learned representations.
• Train per-task binary masks (task masks) to protect units important to dissimilar tasks and to identify useful units for the current task.
• Use task ID embeddings to derive task masks via a sigmoid-based pseudo-gate with a scaling parameter s.
• When similar tasks exist, apply a knowledge transfer attention (KTA) to selectively combine masked outputs from similar tasks for forward transfer.
• Train a separate f_KTA classifier using the transferred knowledge for the current task, enabling forward transfer and allowing backward updates to similar past tasks.
• In parallel, train a reference loss for similarity detection by comparing a transfer model to a no-transfer baseline to determine TSV (task similarity vector).
• Detect similarity online via a transfer model f_k→t and a reference model f_∅, classifying k as similar to t if transfer improves validation loss.

Experimental results

Research questions

• RQ1 Can a single network learn a mixed sequence of similar and dissimilar tasks while avoiding forgetting for dissimilar tasks?
• RQ2 Can the model selectively transfer knowledge from similar tasks to improve new task learning and also permit backward transfer to improve past tasks?
• RQ3 Can the system automatically detect task similarity to trigger appropriate transfer and protection mechanisms?
• RQ4 Does the proposed CAT architecture outperform baselines on mixed-task sequences and demonstrate ablation-supported contributions of its components?

Key findings

• CAT consistently outperforms strong TCL baselines on mixed sequences of similar and dissimilar tasks.
• Forward knowledge transfer is highly effective for similar tasks.
• Backward knowledge transfer can further improve some similar tasks after learning new tasks.
• Knowledge masks effectively prevent forgetting for dissimilar tasks while leaving room for transfer.
• Task similarity detection (TSV) and knowledge transfer attention (KTA) are essential; ablations reduce performance.
• CAT shows robustness across two types of datasets with 2-layer fully connected networks and CNN-based Architectures.