Deep learning ?
Deep learning is part of Machine learning, which is again part of Artificial Intelligence. Currently machine learning is more or less equal to Artificial Intelligence. This has to do with the fact that we are not yet capable of true AI, which means that computers can really think. Being able to break down problems into partial problems and partial solutions into concrete solutions. Basically, machine learning is a form of statistics in which input (data) through various layers (neural networks) examines relationships.
For example, with language it is important how often words occur in combination with each other and how often a structure (grammar) can be discovered. At a higher level what the context of a sentence is. Is it a question and what is the subject, personal form, etc. This is how a computer learns language.
Image recognition
With image recognition, the layers are first pixel patterns, which at a higher level can recognize the edges of a structure. These in turn recognize contours such as a wheel or a nose. On the next level, the contours can be compared to objects such as a car or a human. After that, you can recognize behavior or emotion in a person. For image recognition, so-called convolutional neural networks (CNN) are used, which contain the algorithms to investigate these various layers. In simplified form, a neural network looks like this.
To understand how such a neural network works, let’s take the wheel as an example. A wheel has a fixed shape (round). All pixels of the tires are black. By comparing several wheels and mark them as a wheel, this is called annotating. The pattern is recorded in the relationships between the pixels and the shape. The arrows indicate the relationships and the pixels the dots. At the next wheel, the network will examine whether the relationships and pixels have the same strong relationship. If so the conclusion is “It’s a wheel”. How likely is determined with a percentage of certainty. So the more example wheels, the better the network can determine whether the next wheel also meets them. A neural network is a form of deep learning.
Comparing apples and oranges with deep learning
A nice example of how Deep learning can tackle a problem is how people compare apples and oranges. After all, every person recognizes an apple and an orange. But if it get’s more complex we have significantly more difficulty with it. However, a computer can compare apples and oranges just as easily as a cancer cell to a normal cell. And with high reliability without realizing what one is in relation to the other. The patterns must be so different that the difference is as great or greater than between an apple and a pear. So photos with a very high resolution are needed for cells to be able to determine that distinction. But do not forget also good pathologists who can annotate. If someone does that without knowledge, it becomes a mess.
Statistics
It is important to realize that Machine learning and deep learning use statistics for recognizing patterns. For example, a percentage is assigned to the strength of the relationship for each relationship. The higher that percentage is, the better a sentence can be parsed or an image recognized. That is why the more data the better an AI system becomes. It is also important that the data itself is correct and that there are also differences. For example, if you were to provide 400 photos of mouths and mark them all as mouths, the chances are very high that the next image of a mouth will actually be found. But if that is the only data there is, a nose will also be recognized as a mouth, after all the world consists only of mouths.
Centillien and deep learning
Centillien has already created various products and solutions based on deep learning. For example, Intra, our product for image and recognition of behavior, is based on this. It uses Convolutional Neural Networks to create layers of pattern recognition and then apply deep learning to discover the relationship. The result is astonishing, as the video below makes clear.
What’s in it for me?
Deep learning is a very hot topic at the moment. Understanding how this works doesn’t just help us understand the world a little bit better. It also helps to come up with new applications. So get inspired and come up with an AI application based on deep learning for your organization and ask us to see what we can do with it. Hopefully we will meet soon.
