Mathematica Neural Net Framework

The Mathematica Neural Net Framework is a high-level deep learning framework that is part of the Wolfram Language/Mathematica, rather than being a standalone package. Some nice features of this framework:

• Automatic support variable-length sequences without the need for padding, without sacrificing performance on GPUs. I co-wrote a post for O’Reilly explaining some of the tricks we used to achieve this.
• Uses MXNet as a backend (like how Keras uses TensorFlow). I added a signficant number of features to MXNet to support things like variable-length sequence handling.
• Supports multi-GPU training, so designed to be scalable.
• A belief that having access to an extensive repository of pre-trained nets is essential for a high-level deep learning framework. I lead the creation of the Wolfram Neural Net Repository for this goal.