SparseRecurrentNetwork is an experimental framework for developing and testing deep, recurrent neural networks for sequence prediction. It relies on Tensor Factorization based optimization:
The experimental architecture uses further among others
- sparse autoencoders
- a custom cell unit containing feedforward, recurrent and feedback connections with a custom update logic
- gradient descent with momentum and adaptive gradient updates
- dropout and inhibition
- audio and text input preprocessers
It is written in Python using Google's TensorFlow to provide easy, simple developing-testing cycles (parallelization and cluster deployment is currently WIP as well as development in Scala/Spark). It is part of ReCog Technologies