25 Apr 2021
Analysis of small datasets presents a number of essential challenges not in the least due to insufficient sampling of characteristic patterns in the data making confident conclusions about the unknown distribution elusive and resulting in lower statistical confidence and higher error. In this work, a novel approach to augmentation of small datasets is proposed based on an ensemble of neural network models of unsupervised generative self-learning. Applying generative learning with an ensemble of individual models allowed to identify stable clusters of data points in the latent representations of the observable data. Several techniques of augmentation based on identified latent cluster structure were applied to produce new data points and enhance the dataset. The proposed method can be used with small and extremely small datasets to identify characteristics patterns, augment data and in some cases, improve accuracy of classification in the scenarios with strong deficit of labels.