These priorities have been set by what the Skymind has seen demand for among clients and open-source community members. Contributors are welcome to add features whose priority they deem to be higher.
- CUDA rewrite for ND4J (under way)
- CPU optimizations (C++ backend)
- Hyperparameter optimization (underway, basics done: Arbiter)
- Parameter server
- Sparse support for ND4J
- Performance tests for network training vs. other platforms (and where necessary: optimizations)
- Performance tests for Spark vs. local (ditto)
- Building examples at scale
- OpenCL for ND4J
- CTC RNN (for speech etc.)
Nice to have:
- Automatic differentiation
- Proper complex number support for ND4J (+optimizations)
- Reinforcement learning
- Python support/interface
- Support for ensembles
- Variational autoencoders
- Generative adversarial models
- Hessian free optimization
- Other RNN types: multi-dimensional; attention models, Neural Turing Machine, etc.
- 3D CNNs
This is a work in progress. Last updated Feb. 27 2016.