Compositionality as Learning Bias in Generative RNNs solves the Omniglot Challenge
One aspect of learning to learn concerns the development of compositional knowledge structures that can be flexibly recombined in a semantically meaningful manner to analogically solve related problems. We focus on learning to learn one-shot/few-shot generation and classification tasks of handwritten character trajectories, as described in the Omniglot challenge. We show that solving the challenge becomes possible, by suitably fostering a generative LSTM network to develop well-structured, compositional encodings, which can be quickly reassembled into new, unseen but related character trajectories. This is a major improvement compared to the original approach, which explicitly provided character components. We believe that the development of similarly compressed, compositional structures may also be highly useful to address related learning to learn challenges in other dynamic processing, prediction, and control domains.
PDF Abstract