Self-SLAM: A Self-supervised Learning Based Annotation Method to Reduce Labeling Overhead

Abstract

In recent times, Deep Neural Networks (DNNs) have been effectively used to tackle various tasks such as emotion recognition, activity detection, disease prediction, and surface classification. However, a major challenge in developing models for these tasks requires a large amount of labeled data for accurate predictions. The manual annotation process for a large dataset is expensive, time-consuming, and error-prone. Thus, we present SSLAM (Self-supervised Learning-based Annotation Method) framework to tackle this challenge. SSLAM is a self-supervised deep learning framework designed to generate labels while minimizing the overhead associated with tabular data annotation. SSLAM learns valuable representations from unlabeled data that are applied to the downstream task of label generation by utilizing two pretext tasks with a novel $log-cosh$ loss function. SSLAM outperforms supervised learning and Value Imputation and Mask Estimation (VIME) baselines on two datasets - Continuously Annotated Signals of Emotion (CASE) and wheelchair dataset. The wheelchair dataset is our novel unique surface classification dataset collected using wheelchairs showcasing our framework’s effectiveness in real-world scenarios. All these results reinforce that SSLAM significantly reduces the labeling overhead, especially when there is a vast amount of unlabeled data compared to labeled data. The code for this paper can be viewed at the following link: https://github.com/Alfiya-M-H-Shaikh/SSLAM.git