Abstract
In this paper, we introduce a selective zero-shot classification problem: howcan the classifier avoid making dubious predictions? Existing attribute-basedzero-shot classification methods are shown to work poorly in the selectiveclassification scenario. We argue the under-complete human defined attributevocabulary accounts for the poor performance. We propose a selective zero-shotclassifier based on both the human defined and the automatically discoveredresidual attributes. The proposed classifier is constructed by firstly learningthe defined and the residual attributes jointly. Then the predictions areconducted within the subspace of the defined attributes. Finally, theprediction confidence is measured by both the defined and the residualattributes. Experiments conducted on several benchmarks demonstrate that ourclassifier produces a superior performance to other methods under therisk-coverage trade-off metric.