Towards Information Theoretic Measurement of Fidelity and Diversity in Handwriting Synthesis

摘要

One of the basic goals in handwriting synthesis is to produce samples with both sufficient fidelity (legibility) and diversity. It is not easy in general to automatically evaluate these two qualities, however. In this paper, we show how a relatively simple neural network-based generative model can be used to synthesize online handwritten digits that look quite natural. This autoregressive model contains a parameter which directly controls the fidelity and diversity of generated samples, and must be tuned in order to make a tradeoff between them. We show that optimizing this parameter on the basis of likelihood generally leads to poor results. Then, we propose another objective function for fidelity-diversity tradeoff which is derived from an information theoretic measure, and is approximated via sampling. The proposed approach has been experimented on English and Farsi digits, and as it was observed, the obtained handwriting samples are mostly legible and fairly diverse (representing multiple styles). Also, the plausibility of the proposed performance measure has been compared with log-likelihood and Maximum Mean Discrepancy.