ESL: Explain to Improve Streaming Learning for Transformers

Meghna P. Ayyar, Jenny Benois-Pineau, Akka Zemmari

Abstract: AI systems in real-world scenarios must be able to learn continuously from a stream of data while adapting quickly to concept drift. We propose a training strategy called Explain to improve Streaming Learning (ESL) for the online streaming learning setting where the models have to learn from data on the fly in a single pass. ESL leverages model explanations to identify salient input regions, guiding the streaming learner to focus on these regions by masking the non-salient ones during training. In this work, we focus on using transformers for streaming learning and adapt our CNN-based Feature Explanation Method (FEM) [13] to propose Rollout-FEM with ESL for transformers. We validate the ESL strategy for our streaming learners Entropy-based Move-To-Data (EMTD), its variant EMTD with re-targetting (EMTDR), and the state-of-the-art streaming learning method ExStream [14] and benchmark it on two streaming learning datasets and a real-world egocentric video dataset. Our experiments demonstrate that training with the explanation-based ESL strategy has a better performance than standard training, and EMTDR with ESL achieves the best performance compared to ExStream across the datasets.