Improving Zero-Shot Learning Baselines with Commonsense Knowledge

Zero-shot learning — the problem of training and testing on a completely disjoint set classes relies greatly its ability to transfer knowledge from train test classes. Traditionally semantic embeddings consisting human-defined attributes or distributed word are used facilitate this by improving association between visual embeddings. In paper, we take advantage explicit relations nodes defined in ConceptNet, commonsense graph, generate class labels using graph convolution network-based autoencoder. Our experiments performed three standard benchmark datasets surpass strong baselines when fuse our with existing embeddings, i.e., This work paves path more brain-inspired approaches zero-short learning.