Deep learning with mxnet

Background

Deep neural networks have achieved state-of-the-art results in many tasks and they are showing the power in more and more fields. However, compared with other programming languages, a powerful deep learning tool is still lacking in R.

As a new lightweight and flexible deep learning platform, MxNet provides a portable backend, which can be called from R side.

Related work

There are existing packages on neural networks, like nnet and deepnet.

However, nnet is only working on the single layer network and deepnet doesn't have good performance. Using mxnet, all the computational work can be accelerated by GPUs, which has been one the mainstream method in deep learning.

Details of your coding project

We already have the basic R interface, however, we are still lacking some important models, like LSTM.

Expected impact

Right now there is no other GPU-accelerated deep learning tool in R. We believe this will be a popular choice for many R users.

Mentors

Qiang Kou ([@](mailto:qkou {at} umail {dot} iu {dot} edu)) and Yuan (Terry) Tang ([@](mailto:terrytangyuan {at} gmail {dot} com))

Tests

Fork the package from GitHub and create a pull-request implementing a fix for an open issue ticket of your choice.

Solutions of tests

Students, please post a link to your test results here.

Yun Yan (https://github.com/Puriney/mxnet) with PR #1554 (https://github.com/dmlc/mxnet/pull/1554)