Distill and Replay for Continual Language Learning

Proceedings of the 28th International Conference on Computational Linguistics, pages 3569–3579 Barcelona, Spain (Online), December 8-13, 2020 3569 Distill and Replay for Continual Language Learning Jingyuan Sun,1,2 Shaonan Wang,1,2 Jiajun Zhang,1,2,4 Chengqing Zong1,2,3 1National Laboratory of Pattern Recognition, CASIA, Beijing, China 2University of Chinese Academy of Sciences, Beijing, China 3CAS Center for Excellence in Brain Science and Intelligence Technology, China 4Beijing Academy of Artificial Intelligence, Beijing, China {jingyuan.sun, shaonan.wang, jjzhang, cqzong}@nlpr.ia.ac.cn Abstract Accumulating knowledge to tackle new tasks without necessarily forgetting the old ones is a hall- mark of human-like intelligence. But the current dominant paradigm of machine learning is still to train a model that works well on static datasets. When learning tasks in a stream where data distribution may fluctuate, fitting on new tasks often leads to forgetting on the previous ones. We propose a simple yet effective framework that continually learns natural language understanding tasks with one model. Our framework distills knowledge and replays experience from previ- ous tasks when fitting on a new task, thus named DnR (distill and replay). The framework is based on language models and can be smoothly built with different language model architec- tures. Experimental results demonstrate that DnR outperfoms previous state-of-the-art models in continually learning tasks of the same type but from different domains, as well as tasks of radi- cally different types. With the distillation method, we further show that it’s possible for DnR to incrementally compress the model size while still outperforming most of the baselines. We hope that DnR could promote the empirical application of continual language learning, and contribute to building human-level language intelligence minimally bothered by catastrophic forgetting. 1 Introduction Humans and many advanced animals can learn new tasks without necessarily forgetting the old ones (Glenberg, 1997; Zenke et al., 2017). This ability to continuously learn, accumulate knowledge and reuse them to tackle new challenges through the lifespan is a critical requirement for human-like intelligence. However, the currently dominant paradigm for machine learning is still to train a model on a static dataset, where state-of-the-art methods deliver impressive performance on that particular task. But when learning tasks in a stream where data distribution may shift, the models generally fail to isolate acquired knowledge and forget previously learned tasks. Such phenomenon is known as catastrophic forgetting. There are mainly two stretches of methods in overcoming catastrophic forgetting. One is data-based method that achieves continual learning through reproducing the data distribution of old tasks (Kamra et al., 2017; Chaudhry et al., 2018). Either storing real or generating pseudo examples, the model replays the training data of experienced tasks when learning a new one. The other stretch is model-based that enables continual learning with modification to the model architecture (Schwarz et al., 2018; Masse et al., 2018). For example, some methods (Lee et al., 2017; Aljundi et al., 2018) regularize the loss function to constrain the updates of weights, especially those important for solving tasks. In the continual learning settings of the above methods, the tasks in a stream …

Content truncated to 500 words