Concept splatters: Exploration of latent spaces based on human interpretable concepts

Abstract

Similarity maps show dimensionality-reduced activation vectors of a high number of data points and thereby can help to understand which features a neural network has learned from the data. However, similarity maps have severely limited expressiveness for large datasets with hundreds of thousands of data instances and thousands of labels, such as ImageNet or word2vec. In this work, we present “concept splatters” as a scalable method to interactively explore similarities between data instances as learned by the machine through the lens of human-understandable semantics. Our approach enables interactive exploration of large latent spaces on multiple levels of abstraction. We present a web-based implementation that supports interactive exploration of tens of thousands of word vectors of word2vec and CNN feature vectors of ImageNet. In a qualitative study, users could effectively discover spurious learning strategies of the network, ambiguous labels, and could characterize reasons for potential confusion.