Iterative Weighted Transductive Learning for Handwriting Recognition.

The established paradigm in handwriting recognition techniques involves supervised learning, where training is performed over fully labelled (transcribed) data. In this paper, we propose a weak supervision technique that involves tuning the trained network to perform better on a specific test set, after having completed its training with the standard training data. The proposed technique is based on the notion of a reference model, to which we refer as the “Oracle”, which is used for test data inference and retraining in an iterative manner. As test data that are erroneously labelled will be a hindrance to model retraining, we explore ways to properly weight Oracle labels. The proposed method is shown to improve model performance as much as \(2\%\) for Character Error Rate and \(5\%\) for Word Error Rate. Combined with a competitive convolutional-recurrent architecture, we achieve state-of-the-art recognition results in the IAM and RIMES datasets.