Neural network classification of degraded imagery using soft labels: towards human-level performance with “accurate” likelihoods?

Abstract

The vast majority of recent progress in deep learning for computer vision has been demonstrated on problems with low irreducible error rates. While it is natural to use hard “one-hot” encoded training labels for such problems, this may not be the case in applications with large irreducible error. This includes classification problems on severely degraded or classambiguous imagery. Furthermore, databases primarily consisting of degraded examples are difficult to diagnose in terms of assuring that non-causal statistical correlations across the training and test sets do not exist for certain classes. Expert image analysts however, are typically well-regularized and will not overfit to such correlations. In this work, soft labels are applied to a surrogate problem with large irreducible error, where the labels are generated by an ensemble of networks serving as a proxy for human expert labelers. Results of networks trained on these soft targets, versus their one-hot counterparts are compared. A concept for an “imagination mechanism” in neural networks training on soft labels is also introduced.