- DL4J. DL4J.org. In Java. Commercially supported. Distributed. Integrates with Hadoop/YARN and Spark. Includes a Scala API. Integrates with a vectorization "Rosetta Stone", Canova, which makes it really easy to upload data.
- Keras. Keras.io. In Python. Capable of running on top of either TensorFlow or Theano. Focuses on fast prototyping.
- Caffe. Develped by Yangqing Jia (BVLC: Berkeley Vision and Learning Center). Widely used machine-vision library that ported Matlab’s implementation of fast convolutional nets to C and C++. Not intended for other deep-learning applications such as text, sound or time series data.
- TensorFlow. In Python. Developed by Google. Not distributable.
- Torch7. In Lua. Developed in NYU.
- PyLearn2. Developed by LISA (Dr. Yoshua Bengio's lab). Non-distributed.
- DMTK. Developed by Microsoft.
- Apache Mahout.
Theano, Torch and Caffe employ a BSD License. Deeplearning4j and ND4J are under Apache 2.0 License.
Apache 2.0 License gives resources to defend derivative authors in litigation, and discourage them from attacking others. BSD doesn’t address the issue.
Convolutional networks and Recurrent networks
Multi-input and multi-output training
Quora: what are good open source DL packages?. Recurrent nets (LSTMs), Recursive and stacked denoising autoencoders, Word2vec/Doc2vec/GloVe, RBMs and DBNs, Deep Autoencoders, Convolutional Nets, Recursive Neural Tensor Networks. multilayer Perceptrons, restricted Boltzmann machines, Stacked Denoising Autoencoders and Convolutional nets.