On the Variance of Unbiased Online Recurrent Optimization
This work addresses variance reduction in online gradient-based learning for RNNs, which is an incremental improvement for researchers in recurrent neural networks.
The authors analyzed the variance of the Unbiased Online Recurrent Optimization (UORO) algorithm's gradient estimate for RNNs and proposed modifications to reduce this variance, showing improvements in both theory and practice.
The recently proposed Unbiased Online Recurrent Optimization algorithm (UORO, arXiv:1702.05043) uses an unbiased approximation of RTRL to achieve fully online gradient-based learning in RNNs. In this work we analyze the variance of the gradient estimate computed by UORO, and propose several possible changes to the method which reduce this variance both in theory and practice. We also contribute significantly to the theoretical and intuitive understanding of UORO (and its existing variance reduction technique), and demonstrate a fundamental connection between its gradient estimate and the one that would be computed by REINFORCE if small amounts of noise were added to the RNN's hidden units.