Simplified LSTM unit and search space probability exploration for image description
Oliver Nina, Andrés Rodríguez · 2015
We present a novel method for addressing the semantic description of images. Our method offers two main contributions. First we introduce a recurrent unit that we call a simplified long short-term memory (LSTM) unit which, in contrast to traditional LSTM units, couples the functions of the input and forget gates into a single gate; we observed that this simpler unit improves accuracy. We also propose a novel algorithm for exploring the search space of sentences inferred through a joined Convolutional Neural Network (CNN) and our simplified LSTM unit. We explore the search space by reducing it to a search over sequential trees for the combination of sequences that best represent the image to be described. Our results show improvement over the state of the art methods on the COCO [1] and Flickr8K [2] datasets.