Improving zero-shot learning by mitigating the hubness problem
This addresses a key bottleneck in zero-shot learning for applications like multilingual and image-based systems, though it is incremental as it builds on existing paradigms.
The paper tackled the hubness problem in zero-shot learning, where certain vectors dominate neighborhoods and degrade label accuracy, and proposed a correction method that improved performance across cross-lingual, image labeling, and image retrieval tasks.
The zero-shot paradigm exploits vector-based word representations extracted from text corpora with unsupervised methods to learn general mapping functions from other feature spaces onto word space, where the words associated to the nearest neighbours of the mapped vectors are used as their linguistic labels. We show that the neighbourhoods of the mapped elements are strongly polluted by hubs, vectors that tend to be near a high proportion of items, pushing their correct labels down the neighbour list. After illustrating the problem empirically, we propose a simple method to correct it by taking the proximity distribution of potential neighbours across many mapped vectors into account. We show that this correction leads to consistent improvements in realistic zero-shot experiments in the cross-lingual, image labeling and image retrieval domains.