Source-linked AI summary
Continual Learning of a Mixed Sequence of Similar and Dissimilar Tasks
Zixuan Ke, Bing Liu, Xingchang Huang
TL;DR
Continual learning must preserve prior knowledge while transferring knowledge across similar tasks, yet existing methods do not jointly address these goals for mixed similar and dissimilar sequences. CAT uses a shared knowledge base, task similarity detection, task masks, and selective attention, and experiments show that it outperforms strong baselines across mixed-task evaluations.
Problem
Existing continual-learning methods do not provide forgetting avoidance together with forward and backward knowledge transfer for mixed sequences of similar and dissimilar tasks.
Method
CAT detects similar previous tasks, protects dissimilar-task knowledge with task masks, and selectively transfers and updates similar-task knowledge through knowledge transfer attention in a shared network.
Results
CAT outperforms strong baselines overall, performs better on similar tasks, handles forgetting reasonably well on dissimilar tasks, and achieves effective forward and backward transfer.
Takeaways & Limitations
The results support CAT as a continual-learning architecture for mixed task sequences that combines forgetting avoidance with selective knowledge transfer.
Abstract
from arXiv · showhide
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.
1 Introduction
Continual learning aims to accumulate knowledge while avoiding forgetting and transferring knowledge forward and backward across similar tasks. CAT addresses these goals for mixed sequences of similar and dissimilar classification tasks.
- Motivation: Continual learning should avoid forgetting previous tasks while transferring knowledge forward and backward when tasks share knowledge.The paper identifies these capabilities as simultaneous goals for continual learning.
- Motivation: In practical sequences, a new task may be similar to some previous tasks and dissimilar to others.The learner must preserve prior knowledge while using shared knowledge from similar tasks.
- Research gap: Most existing continual-learning models focus on catastrophic forgetting, while parameter updates can degrade models for previous tasks.The cited prior work primarily addresses forgetting rather than simultaneous transfer across mixed tasks.
- CAT model: CAT uses a shared knowledge base, detects similar previous tasks, protects dissimilar-task knowledge with task masks, and selectively transfers similar-task knowledge with attention.It also learns masks for the new task and updates transferable past knowledge to support backward transfer.
- Contribution: CAT is proposed to learn mixed sequences of similar and dissimilar tasks while achieving forgetting avoidance and forward and backward knowledge transfer.The paper presents this as a generalization of existing task continual-learning work.
2 Related Work
Prior continual-learning research largely emphasizes catastrophic-forgetting mitigation, while other methods address transfer or task-specific settings separately. The paper positions CAT as combining forgetting avoidance and knowledge transfer for mixed task sequences.
- Forgetting mitigation: Continual-learning research includes distillation, weight-importance estimation, parameter generation, orthogonal updates, and loss-based approaches to mitigate forgetting.These methods constrain or adapt updates so new-task learning damages prior-task models less.
- Task continual learning: In task continual learning, methods such as GEM, A-GEM, and HAT use task identities, while HAT protects old models with hard attention masks.The paper distinguishes these approaches from CAT because they do not provide knowledge-transfer mechanisms.
- Research gap: Existing task-continual-learning methods do not simultaneously handle forgetting and transfer knowledge to improve new-task learning.The paper identifies this combination as a gap in the related work.
- Knowledge transfer: Progressive Networks provide forward transfer but cannot perform backward transfer, grow quadratically with task count, and do not handle mixed task sequences.Their separate-model construction creates scalability and task-sequence limitations.
- Lifelong learning: Earlier lifelong-learning studies also explored forward and backward transfer, but mainly used traditional methods such as regression, naive Bayes, and KNN.The paper contrasts those methods with its neural-network task-continual-learning setting.
3 Proposed CAT Model
CAT learns mixed sequences of similar and dissimilar tasks in one network by protecting units used by dissimilar tasks and selectively transferring knowledge from similar tasks. It detects task similarity and uses task masks to support both forward and backward transfer.
- Architecture and task masks: CAT uses a shared knowledge base and task-specific masks to learn mixed sequences of similar and dissimilar tasks.The masks identify units important for each task and protect them during future learning.
- Preventing forgetting: For dissimilar previous tasks, CAT blocks gradients through units marked as used, preventing updates from overwriting their learned knowledge.Used units are accumulated across dissimilar tasks, and corresponding gradients are set to zero while learning the new task.
- Architecture and task masks: Task masks are learned from task ID embeddings and indicate which knowledge-base units are important for each task.The masks are trained with the task classifier, using a hard-attention-inspired mechanism and annealing toward binary values.
- Knowledge transfer: For similar tasks, knowledge transfer attention assigns different weights to previous task representations and combines them for the new task.The transferred units remain updateable, allowing backward knowledge transfer to improve similar previous task models during new-task training.
- Task similarity detection: CAT detects task similarity by comparing a transfer model with a reference model trained independently for the new task.A previous task is labeled similar when its transfer model performs better on the new task’s validation data than the reference model.
4 Experiments
CAT is evaluated on mixed sequences combining similar and dissimilar tasks, using multiple dataset constructions, baselines, architectures, and ablations. Across these experiments, CAT achieves strong overall performance by combining forgetting avoidance with selective knowledge transfer.
- Datasets: The evaluation constructs four mixed-sequence datasets from similar-task federated-learning data and dissimilar-task EMNIST or CIFAR100 data.The mixed datasets combine EMNIST-10 or EMNIST-20 with F-EMNIST, and CIFAR100-10 or CIFAR100-20 with F-CelebA.
- Compared Baselines: CAT is compared with ten task continual learning baselines, including methods targeting forgetting, forward transfer, parameter masking, replay, and knowledge sharing.The compared systems include EWC, HAT, UCL, HYP, HYP-R, PRO, PathNet, and RPSNet, among others.
- Overall Performance: CAT outperforms all baselines in overall accuracy across the four mixed sequence datasets.Forgetting-oriented methods outperform NCL but remain significantly worse than CAT, while several transfer-oriented methods fail to outperform NCL on mixed sequences.
- Dissimilar Tasks: CAT performs better than most baselines on dissimilar tasks while dealing with forgetting reasonably well.CAT performs similarly to HAT, whereas PathNet and RPSNet vary sharply between EMNIST-20 and F-EMNIST.
- Similar Tasks: CAT markedly outperforms HAT and other baselines on similar tasks by leveraging shared knowledge among them.Other continual learning approaches, including HAT, primarily avoid interference with important units from previous tasks.
- Knowledge Transfer: Forward knowledge transfer is highly effective, while backward transfer slightly improves F-MNIST performance and markedly improves F-CelebA performance.Forward transfer uses accuracy when each similar task is first learned; backward transfer uses the final result after all tasks are learned.
- Ablation Study: The full CAT system gives the best overall accuracy in every ablation comparison, with both knowledge-transfer attention and task-similarity detection contributing to performance.The ablations remove KTA or task-similarity-vector detection, including settings that treat all previous tasks as dissimilar or similar.
5 Conclusion
The paper presents CAT as a continual-learning architecture intended to combine forgetting avoidance, forward and backward knowledge transfer, and learning mixed sequences of similar and dissimilar tasks. Experiments show that CAT outperforms strong baselines, while future work targets similarity estimation, efficiency, and reduced labeling.
- Contributions: CAT targets four capabilities: no forgetting, forward knowledge transfer, backward knowledge transfer, and learning mixed sequences of similar and dissimilar tasks.The paper states that no existing continual-learning method had all four capabilities.
- Results: Experimental results show that CAT outperforms strong baselines.
- Future Work: Future work includes improving similar-task accuracy, removing explicit similarity computation, and using fewer labeled data in training.The paper specifically mentions regression-based task-similarity computation as one possible direction.
Broader Impact
The paper motivates continual learning by the growing need for AI agents to learn many skills, some similar and some distinct. It argues that such agents should avoid interference while benefiting from shared knowledge.
- Motivation: AI agents such as chatbots, personal assistants, and physical robots increasingly need to learn many skills or tasks.
- Motivation: Because tasks may be similar or distinct, agents should learn without interference while improving through shared transferable knowledge.