Training deep nets with imbalanced and unlabeled data

Training deep belief networks (DBNs) is normally done with large data sets. Our goal is to predict traces of the surface of the tongue in ultrasound images of human speech. Hand-tracing is labor-intensive; the dataset is highly imbalanced since many images are extremely similar. We propose a bootstrapping method which handles this imbalance by iteratively selecting a small subset of images to be hand-traced (thereby reducing human labor time), then (re)training the DBN, making use of an entropy-based diversity measure for the initial selection, thereby achieving over a two-fold reduction in human time required for tracing with human-level accuracy.