A Framework and Benchmark for Deep Batch Active Learning for Regression
David Holzmüller, Viktor Zaverkin, Johannes Kästner, Ingo Steinwart; 24(164):1−81, 2023.
Abstract
This study focuses on active learning methods for improving the sample efficiency of neural network regression by adaptively selecting batches of unlabeled data for labeling. The authors present a framework that utilizes base kernels, kernel transformations, and selection methods to construct these active learning methods. The framework encompasses Bayesian methods based on Gaussian process approximations of neural networks as well as non-Bayesian methods. Additionally, the authors propose using sketched finite-width neural tangent kernels instead of the commonly used last-layer features, and combining them with a novel clustering method. To evaluate the effectiveness of different methods, the authors introduce an open-source benchmark consisting of 15 large tabular regression datasets. The proposed method outperforms the state-of-the-art on the benchmark, can scale to large datasets, and does not require adjustments to the network architecture or training code. The authors provide open-source code that includes efficient implementations of all kernels, kernel transformations, and selection methods, allowing for reproduction of their results.
[abs]
[code]