Understanding Convolutional Neural Networks in Terms of Category-Level Attributes

摘要

It has been recently reported that convolutional neural networks (CNNs) show good performances in many image recognition tasks. They significantly outperform the previous approaches that are not based on neural networks particularly for object category recognition. These performances are arguably owing to their ability of discovering better image features for recognition tasks through learning, resulting in the acquisition of better internal representations of the inputs. However, in spite of the good performances, it remains an open question why CNNs work so well and/or how they can learn such good representations. In this study, we conjecture that the learned representation can be interpreted as category-level attributes that have good properties. We conducted several experiments by using the dataset AwA (Animals with Attributes) and a CNN trained for ILSVRC-2012 in a fully supervised setting to examine this conjecture. We report that there exist units in the CNN that can predict some of the 85 semantic attributes fairly accurately, along with a detailed observation that this is true only for visual attributes and not for non-visual ones. It is more natural to think that the CNN may discover not only semantic attributes but non-semantic ones (or ones that are difficult to represent as a word). To explore this possibility, we perform zero-shot learning by regarding the activation pattern of upper layers as attributes describing the categories. The result shows that it outperforms the state-of-the-art with a significant margin.