We used basic neural networks in Chapter 8, Beating CAPTCHAs with Neural Networks. A recent flood of research in the area has led to a number of significant advances to that base design. Today, research in neural networks is creating some of the most advanced and accurate classification algorithms in many areas.
These advances have come on the back of improvements in computational power, allowing us to train larger and more complex networks. However, the advances are much more than simply throwing more computational power at the problem. New algorithms and layer types have drastically improved performance, outside computational power.
In this chapter, we will look at determining what object is represented in an image. The pixel values will be used as input, and the neural network will then automatically find useful combinations of pixels to form higher-level features. These will then be used for the actual classification. Overall, in this chapter, we will examine the following:
- Classifying objects in images
- The different types of deep neural networks
- Theano, Lasagne, and Nolearn; libraries to build and train neural networks
- Using a GPU to improve the speed of the algorithms
Computer vision is becoming an important part of future technology. For example, we will have access to self-driving cars in the next five years (possibly much sooner, if some rumors are to be believed). In order to achieve this, the car's computer needs to be able to see around it: obstacles, other traffic, and weather conditions.
While we can easily detect whether there is an obstacle, for example using radar, it is also important we know what that object is. If it is an animal, it may move out of the way; if it is a building, it won't move at all and we need to go around it.