LPC: A Logits and Parameter Calibration Framework on Continual Learning

Anonymous

Abstract: Deep learning based pre-trained natural language processing (NLP) models typically pre-train on large unlabeled corpora first, then fine-tune on new tasks. When we execute such a paradigm on continuously sequential tasks, the model will suffer from the catastrophic forgetting problem (i.e., they forget the parameters learned in previous tasks when we train the model on newly emerged tasks). Inspired by the idea of how humans learn things, we aim to maintain the old knowledge when we transfer to novel contents and calibrate the old and new knowledge. We propose a Logits and Parameter Calibration (LPC) framework to reduce the catastrophic forgetting in the continual learning process. The proposed framework includes two important components, the Logits Calibration (LC) and Parameter Calibration (PC). The core idea is to reduce the difference between old knowledge and new knowledge by doing calibration on logits and parameters so that the model can maintain old knowledge while learning new tasks without preserving data in previous tasks. First, we preserve the parameters learned from the base tasks. Second, we train the existing model on novel tasks and estimate the difference between base logits and parameters and novel logits and parameters. Third, we drift from the base tasks to novel tasks gradually. Furthermore, we integrate the logtis and parameter calibration into a brand-new optimization algorithm. Finally, we do experiments on 7 scenarios of the GLUE (the General Language Understanding Evaluation) benchmark. The experimental results show that our model achieves state-of-the-art performance on all 7 scenarios.

Paper Type: long

Consent To Share Data: yes