Zero-shot Active Learning with Topological Clustering for Multiclass Classification

Anonymous

Keywords: Topological clustering, active learning, multi-class classification.

TL;DR: This paper tackles the problem of reducing the annotation burden with topological persistance for multi-class classification.

Abstract: We present a novel approach for zero-shot active learning for multi-class classification based on a clustering technique, called ToMATo, which is guided by topological persistence. Our objective is to identify effective regions in the feature space for label querying. The labeling of examples in these regions will allow the training of efficient multi-class classification prediction functions. We have adapted ToMATo with a density aware $\delta$-Rips graph in order to obtain homogeneous simplicial trees. From these trees, informative simplices are identified with respect to the annotation effort, or the budget. Representative examples from each of them are labeled by an oracle and these labels are then propagated through the trees. We adapt ToMATo by computing our persistence diagram (PD) from a $\delta$-Rips graph that is estimated using a $k$-nearest neighbor distance matrix. This allows the application of the method to large scale scenarios. From this perspective we also propose a local density estimator from the same distance matrix. Comparisons on different benchmarks show that the proposed approach greatly improves performance with respect to a random querying strategy for label assignment that has been found outperforming state-of-the art approaches in previous works.

Previous Submission: No