Supervised Deep / Recurrent Nets
Supervised artificial recurrent neural networks with adaptive feedback connections are more powerful in sequential data processing than classical statistical methods such as Hidden Markov Models or Condtional Markov Fields. Our deep feed-forward networks also set records in many real-world pattern recognition benchmarks. The goal of this project is to analyze and further improve state-of-the-art algorithms for training recurrent and deep neural networks, also exploring how much they can benefit from massively parallel GPU architectures. The networks will be applied to high-dimensional problems (e.g., object detection) and challenging real-world benchmarks, such as prediction/compression, handwriting recognition, image recognition.