In this section, we will present a number of use cases for mobile deep learning. This is a very different situation from the desktop or cloud deep learning where GPUs and electricity are commonly available. In fact, on a mobile device, it is very important to preserve the battery and GPUs are frequently not available. However, deep learning can be very useful in a number of situations. Let's review them:

• Image recognition: Modern phones have powerful cameras and users are keen to try effects on images and pictures. Frequently, it is also important to understand what is in the pictures, and there are multiple pre-trained models that can be adapted for this, as discussed in the chapters dedicated to CNNs. A good example of a model used for image recognition is given at https://github.com/TensorFlow/models/tree/master/official/resnet.

• Object localization: Identifying moving objects is a key operation and is required for video and image processing. One can, for instance, imagine that if multiple persons are recognized in an image, then the camera will use multiple focus points. A collection of examples for object localization is given at https://github.com/TensorFlow/models/tree/master/research/object_detection.
• Optical character recognition: Recognizing handwritten characters is fundamental in for many activities such as text classification and recommendation. Deep learning can provide fundamental help for carrying out these activities. We looked at a few examples of MNIST recognition in the chapters dedicated to CNNs. Information on MNIST can also be found at https://github.com/TensorFlow/models/tree/master/official/mnist.
• Speech recognition: Voice recognition is a common interface for accessing modern phones. Deep learning is therefore used to recognize voices and spoken commands. The progress made in this area over the last few years has been impressive.
• Translation: Dealing with multiple languages is part of the modern multicultural world. Phones are becoming more and more accurate for on-the-fly translations across languages, and deep learning has helped to break barriers, which was impossible to even imagine a few years ago. We looked at some examples of machine translation during the chapter dedicated to RNNs.
• Gesture recognition: Phones are starting to use gestures as interfaces for receiving commands. Of course, there are models for this.
• Compression: Compression is a key aspect in mobile phones. As you can imagine, it is beneficial to reduce the space before sending an image or a video over the network. Similarly, it might be convenient to compress data before storing locally on the device. In all these situations, deep learning can help. A model using RNNSs for compression is located at https://github.com/TensorFlow/models/tree/master/research/compression.