Correction Networks: Meta-Learning for Zero-Shot Learning

We propose a model that learns to perform zero-shot classification using a meta-learner that is trained to produce a correction to the output of a previously trained learner. The model consists of two modules: a task module that supplies an initial prediction, and a correction module that updates the initial prediction. The task module is the learner and the correction module is the meta-learner. The correction module is trained in an episodic approach whereby many different task modules are trained on various subsets of the total training data, with the rest being used as unseen data for the correction module. The correction module takes as input a representation of the task module's training data so that the predicted correction is a function of the task module's training data. The correction module is trained to update the task module's prediction to be closer to the target value. This approach leads to state-of-the-art performance for zero-shot classification on natural language class descriptions on the CUB and NAB datasets.