Exploring Feature Collocation for Semantic Concept Identification

Abstract

Motivated by object representation in psychology, we present a binary feature classifier for the purpose of semantic concept category identification/classification by incorporating feature distribution. We propose the classification algorithm based on the variant of L1 norm regularized sparse classifiers, where the features are weighted according to their distribution, which is estimated by “maximum collocation”. This method achieves high accuracy in identifying semantic concepts, outperforming standard benchmark methods on a large database of animal and artifact features.


Back to Table of Contents