Are neural networks dependent on machine learning?

Science Year 2019: »Artificial Intelligence« The working worlds of the future will be decisively shaped by the development of artificial intelligence. The Science Year 2019 will deal with the opportunities and challenges of this technology.

In our first blog post we showed that it is not that easy to grasp the term »Artificial Intelligence« and how we nevertheless managed to come to a common definition in the SmartAIwork project. Artificial intelligence is a collective term, in some cases from very different disciplines and varieties, which we want to take a closer look at in today's blog post: machine learning, neural networks and deep learning.

From a historical perspective, the field of artificial intelligence is very broad. The term »machine learning« is used very often in the media. This term is sometimes used synonymously with "artificial intelligence". In the real sense, machine learning is only one of many sub-areas of AI. Machine learning is a class of algorithms and processes that focus on learning from data or empirical knowledge.

Classification of terms in the context of artificial intelligence.

How does this work? Systems train themselves using machine learning processes, e.g. B. via neural networks, themselves certain skills by using known patterns to interpret new decision cases. For this purpose, training data are necessary for the respective decision case, which already anticipate the result. The higher the data quality and the more data there is, the better the system can be trained. By comparing the calculated result with the desired target result, it learns to make better and better decisions.

In the case of an image classifier, for example, the AI ​​would learn to distinguish dogs from cats. For this learning process, the AI ​​must first be given tens of thousands of classified images of dogs and cats as training data. The result is already known in the training set - whether it is a cat or a dog. The system trains, compares its results with the target result, learns from mistakes and can thus improve.

Representation of a neural network using the example of image recognition of dogs and cats.

Another example would be an expert system for automated procurement, e.g. B. of office supplies. The system learns here, on the basis of previous decisions, how optimal procurement should be made. Has a procurement been billed automatically or were manual corrections required by an employee? If a corresponding set of training data is available, the AI ​​can be trained on it and thus learn to make decisions more and more precisely.

Neural networks: layer by layer for the best result

In the large bouquet of machine learning processes, there is a class of learning processes that are currently attracting a great deal of attention: "Artificial Neural Networks". A neural network consists of a large number of interconnected artificial neurons (shown in the picture as blue circles). These are usually arranged in layers. The layer that receives the input data is called the input layer. In the case of image recognition, these would be the individual color points of an image (pixels). The Output Layer generates the output and can, for example, indicate whether the picture shows a cat or a dog. There can be virtually any number of other layers - so-called hidden layers - between the input and output layers. Each neuron now generates an individual output, which is dependent on the respective input and the internal activation function of the neuron. All outputs of a layer are forwarded as inputs for the next layer, with each connection being able to forward the information to different degrees. The strength of each individual connection is determined by a special learning algorithm (back propagation) when creating and training the neural network.

Deep learning: the supreme discipline of artificial intelligence

The neural networks also come in numerous forms. Often it is networks with a large number of "hidden layers" that achieve particularly good results. Such networks are also called deep neural networks, which is why they are summarized under the term "deep learning". Furthermore, there are also networks with feedback to previous layers (recurrent neural networks) and networks that contain special layers for image processing (convolutional neural networks).

Today's AI applications mostly serve very specific purposes. These include, for example, problems such as text understanding, image recognition, expert systems or digital assistants such as chatbots. The applications described are examples from what is known as »weak AI«, i. H. very narrowly defined tasks that an AI can perform self-optimizing if necessary. The »strong AI« or also called super intelligence is not yet feasible today, since the currently known and used processes as well as the current technical possibilities are not sufficient to achieve or even surpass the entire intellectual abilities of people. Nevertheless, processes such as neural networks already enable AI applications to cope with tasks that seemed unthinkable just a few years ago.

Would you like to learn more about how machine learning processes work? In our data science and AI seminars, we introduce you to the field in an easy-to-understand manner and show you which data you need to work with AI yourself and how you can find and process this data in the company.

Reading links: