Gluon: High-Performance Machine Learning Interface

Gluon: High-Performance Machine Learning Interface

AWS and Microsoft have just announced the launch of Gluon, an open-source API for deep learning developers to rapidly and efficiently create high-performance machine learning models.

Gluon’s API contains a library of ready-made and enhanced neural network modules. These will help both beginners and seasoned developers alike to quickly build prototypes and use dynamic neural network graphs for their own creations—with no slowdown on training the network.

Machine learning is an intensive process that may take days to complete. It requires three fundamental parts: a neural network model, data to train with, and the training algorithm for the model. The inputted data is linked to an output (or forecasts) through a series of layers and weights, which are managed by the algorithm. The algorithm adjusts the weights in the network based on the errors in the output.

The time-consuming part of this process comes from defining the network graph at the start, followed by freezing it and adjusting the weights as necessary. Network definition alone may require millions of connections which need to be manually created, making it difficult to test and re-use them. But Gluon avoids those pitfalls.

Gluon features include:

  1. User-friendly interface – easy for new developers to understand.
  2. Dynamic networks – network definition is flexible, thereby easier to manage. Developers may combine elements of fast symbolic representations and dynamic imperative definitions of the network and its algorithms.
  3. Algorithm can dynamically alter network – rather than having two different definitions, the algorithm may change the network during definition and training. This lets developers use advanced algorithms and models that are easy to encode, implement, and test.
  4. High-Performance Training Operators – Due to Gluon’s ability to marry a concise API with formal definition without manual optimization or developer intervention, there is no slowdown in training speed as opposed to other methods.

Gluon is currently available in Apache MXNet and will be open to more frameworks in the future. It’s available today.

Need assistance in getting your business on the cloud? Just contact our AWS experts at PolarSeven.