Zero-shot Learning With Fuzzy Attribute

As the zero-shot problem was proposed in machine learning field, attributes became the key point to solve zero-shot problems. The wildly used binary attribute in zero-shot learning has many limitations, and many researches had made an improvement on it. In this paper, we propose fuzzy attributes, which can describe objects better than binary attributes. We design a classifier to train the fuzzy attributes, and also consider the distance affect attribute in feature space. At last, we take experiment on AwA dataset, and the experimental results shows the fuzzy attribute can play a better performance than binary attributes in zero-shot learning.