Combinets: Learning New Models via Recombination

Modern machine learning methods struggle with problems with small amounts of training data. One solution to is to reuse existing data through transfer methods such as one-shot or transfer learning. However these approaches tend to require an explicit hand-authored or learned definition of how reuse can occur. We present a new representation called conceptual expansions that serves as a general representation for reuse from existing machine-learned knowledge. We evaluate our approach by building conceptual expansions for image classifiers and Generative Adversarial Networks for new classes with as few as ten samples.