Generating Sequences With Recurrent Neural Networks

This technology leverages Long Short-Term Memory (LSTM) recurrent neural networks to generate complex, realistic sequences with long-range structure, predicting one data point at a time.

Generating Sequences with Recurrent Neural Networks, a foundational 2013 paper by Alex Graves, demonstrates how LSTM RNNs can generate intricate sequences by predicting individual data points. The method excels with both discrete data (text) and real-valued data (online handwriting), even extending to handwriting synthesis conditioned on text input. This approach, which models the probability distribution of future events given past information, produces highly realistic cursive handwriting across various styles, showcasing the power of RNNs to capture long-term dependencies and generate novel, plausible sequences.

https://arxiv.org/pdf/1308.0850.pdf