Abstract
We present a novel unsupervised data analysis method, Multi-feature Information Bottleneck (MfIB), which is an extension of the Information Bottleneck (IB). In comparison with the original IB, the proposed MfIB method can analyze the data simultaneously from multiple feature variables, which characterize the data from multiple cues. To verify the effectiveness of MfIB, we apply the corresponding MfIB algorithm to unsupervised image categorization. In our experiments, by taking into account multiple types of features, such as local shape, color and texture, the MfIB algorithm is found to be consistently superior to the original IB algorithm which takes only one source of features into consideration. Besides, the performance of MfIB algorithm is also superior to the state-ofthe-art unsupervised image categorization methods.