Improving Performance of Convolutional Neural Networks via Feature Embedding

Torumoy Ghoshal, Silu Zhang, Xin Dang, Dawn Wilkins, Yixin Chen

Abstract: Recently convolutional neural networks (CNN) have shown exceptional performance with data where a feature structure is explicitly defined, for example image data. Real world data is often represented as d dimensional vectors and they lack such feature structure. If features could be embedded into a low dimensional space to introduce feature locality, CNNs could take advantage of the newly introduced feature structure and show better performance. In this paper, we present a technique of feature embedding to introduce feature locality so that non-image data exhibit image like feature structure. We achieve this by embedding features into a 1d or 2d space using t-SNE. We show that CNN performs better under the proposed approach.