In a recent announcement, Microsoft and Amazon announced Gluon, a new open source deep learning interface that enables developers to create efficient machine learning models and enabling them to do so in a way that is quicker and easier than previously possible.
The Gluon interface provides developers with an API that enables access to a collection of pre-built neural network components that are optimized for efficiency right out of the box. This interface can be especially helpful to developers that are new to machine learning. It enables them to define and manipulate various models in a manner that is similar to traditional object-oriented code. For those developers who are more experienced with machine learning, the interface can also be quite useful. It enables them to create new prototypes. It is also useful for utilizing dynamic neural network graphs without having to put great amounts of effort into learning how to use new tools.
When it comes to working with neural networks and deep learning, Gluon can prove quite useful. The interface includes several innovative features:
- An Easy to Use API — Because Gluon enables developers to build and work with neural networks using code that is clear and easy to learn, it can make it easier to get started than in the past. Previously, developers would have to manually weight and score nodes.
- Dynamic Networks — Unlike more traditional neural networks, Gluon enables networks to scale dynamically, which makes them easier to manage. Developers can also choose to switch back and forth between traditional symbolic representations of neural networks and algorithms and the easier-to-use dynamic definitions.
- Algorithm Defined Networks — With algorithm defined networks, algorithms can be used to automatically modify the network, making it easy to create algorithms and models that were previously more difficult to do. Researchers can now define and use even more complicated algorithms and networks than before, taking advantage of the fact that they can use standard loops and conditional statements when working with them.
- High-Performance Operators for Network Training — In many cases, training a neural network can require a significant amount of processing capability, especially if developers are working with a concise API and are using dynamic network graphs. However, Gluon makes it easy to have both a concise API and dynamic graphs without having to sacrifice network training speed.
For those looking to give Gluon a try, it is currently available in Apache MXNet on AWS, with additional support for the Microsoft Cognitive Toolkit coming in a future release. There are 50 examples available in the AWS Deep Learning Amazon Machine Image (AMI) for those who would like to learn more about Amazon and Microsoft’s Gluon service.
While neural networks can be quite useful when it comes to deep learning and other machine learning tasks, they have traditionally been quite resource-intensive. This can take a great deal of time to configure, train, and optimize. On top of that, finding and resolving bugs within a neural network can be a painstaking process, especially with larger networks. However, services like Gluon can make this process easier for those who are just starting out with neural networks and even for those who have a great deal of experience working with them.