Keras needs a lower-level deep learning library to perform tensor-level manipulations. This lower level deep-learning library is called the backend engine. Possible backend engines for Keras include the following:

• TensorFlow (www.tensorflow.org): This is the most popular framework of its kind and is open sourced by Google.
• Theona (deeplearning.net/software/theano): This was developed at the MILA lab at Université de Montréal.
• Microsoft Cognitive Toolkit (CNTK): This was developed by Microsoft.

The format of this modular deep learning technology stack is shown in the following diagram:

The advantage of this modular deep learning architecture is that the backend of Keras can be changed without rewriting any code. For example, if we find TensorFlow better than Theona for a particular task, we can simply change the backend to TensorFlow without rewriting any code.