- Abstract: Recurrent neural networks (RNNs) are powerful models for sequential data. They can approximate arbitrary computations, and have been used successfully in domains such as text and speech. However, the flexibility of RNNs makes them susceptible to overfitting and regularization is important. We develop a noise-based regularization method for RNNs. The idea is simple and easy to implement: we inject noise in the hidden units of the RNN and then maximize the original RNN's likelihood averaged over the injected noise. On a language modeling benchmark, our method achieves better performance than the deterministic RNN and the variational dropout.