Novel Representative Sampling for Improved Active Learning. Issue 20 (2022)

• Record Type: Journal Article
• Authors: Sarkar, Debangsha; Shabani, Amir; Narayan, Apurva
• Is Part Of: IFAC-PapersOnLine. Volume 55:Issue 20(2022)
• Page Start: 55
• Page End: 60
• Publication Date: 2022
• Subjects: Active learning -- Representative sampling -- Sigma points sampling -- Unscented Kalman filter; Automatic control -- Periodicals; 629.805
• DOI: 10.1016/j.ifacol.2022.09.071
• Languages: English
• ISSNs: 2405-8963

Abstract

Abstract: Active learning solves machine learning problems where acquiring labels for the data is costly. It allows for the learner to select training samples by asking intelligent questions. Various sampling strategies exist for choosing the training set for pool-based active learning. However, the existing representative querying approaches for active learning do not attempt to capture the underlying data distribution, which we believe is an important part of representative sampling. To that end, we propose an adaptation of the sigma point sampling technique from unscented transformation (UT) for constructing a representative subset. UT has shown to be very effective in non-linear transformation modeling in object tracking and robotics. When combined with the Gaussian mixture model, sigma points can estimate the statistical moments such as mean and co-variance of an unknown distribution with very few samples which are generated deterministically. Sigma point sampling being parameterized gives better control over the sampling process. We use sigma points for representative subset construction and train the learner on them. We compare our results with other sampling techniques and improve test accuracy on the handwritten digit recognition data set MNIST.