Characterizing Task Difficulty Using Spatial Entropy

Anonymous

Keywords: meta-learning, meta-features, spatial entropy, meta-knowledge, deep neural networks, learning to learn

TL;DR: Understanding task difficulty using spatial entropy

Abstract: Our study highlights the use of $\textit{spatial entropy}$ as a means to characterize the difficulty of learning tasks. We show how the mutual information of class co-occurrences with regions in the feature space provides an informative curve profile to estimate the degree of difficulty in classification tasks. Empirical results demonstrate the feasibility of employing spatial entropy to quantify the quality of new representations in deep neural networks; results show how spatial entropy can act as a powerful meta-feature to enrich the current family of dataset characterizations.