Few-Shot Learning Text Classification in Federated Environments
Nehal Muthukumar · 2021 Smart Technologies, Communication and Robotics (STCR) · 2021
With increasing privacy restrictions on personal devices, data is often prevented from leaving these devices and reaching the central servers where critical operations are carried out to research and train models better. It is crucial to improve such data and fine-tune the performance of machine learning models in use today. Here, few-shot learning is applicable with federated datasets, where each device holds a limited amount of data only. In this paper, we carry out experiments using an Induction Network (using meta-learning for classification tasks) or a pre-trained BART model by customising it for each node in our dataset. The former helps with a solution devoid of recurrent communication among the node and also slashes computational costs significantly. The latter is bias-free and works efficiently with sparse data. Furthermore, upon reviewing the literature, it was observed that there is a lack of research work entailing the application of few-shot learning algorithms coupled on natural language processing tasks in a federated setting. We attempt to address this gap with this research.