Neural networks (also: artificial neural networks, ANN) represent an important branch of development in the field of artificial intelligence and are the particular subject of research in what is often referred to as neuro-informatics, a relatively new interdisciplinary field between computer science and neurobiology.
Neural networks enable a computer to perform highly efficient pattern recognition in big data using specially-developed algorithms and deep learning. Neural networks can therefore recognise patterns in large amounts of data as a multitude of algorithms in a way that is vaguely modelled after the human brain.
Closely interwoven with the term neural networks is deep learning, which can be described as a process in which neural networks are layered on top of each other in several levels. This is why the term “stacked neural networks” is often used synonymously for deep learning methods. The layer structure makes it possible to analyse or cluster increasingly complex information in large data sets, e.g. to automatically recognise the content of an image or a face – but more about this later.
Although its name suggests otherwise, the structure of neural networks is, strictly speaking, not (yet) completely equal to its biological counterpart, the neurons in the nervous system of a living being. Rather, the metaphor uses the clarification of how different (artificial or natural) neurons work together in combination. However, both the depth structure and the functioning of the networks differ from each other to date: if computer systems are currently still working sequentially to a large extent, biological neural networks are capable of the parallel processing of massive amounts of information.
Neural networks interpret large amounts of (usually unlabelled) raw data through a specific form of machine classification: due to its multi-layered structure, the neural network first identifies basic patterns from training data and then independently derives even more complex patterns from them. The special feature: analogously to the biological nervous system, neuronal networks are able to change the strength of the connections between each other precisely, depending on the importance of the information they are supposed to convey.
If, for example, a face in a photo is to be analysed, the first layer of the arithmetic node would only register differences in the brightness values of the individual pixels. The second layer, on the other hand, would detect vertical and horizontal lines, with some pixels connected to each other, while the third layer would detect shapes and patterns, and so on. In this way, the system can learn on the basis of massive amounts of images that, for example, the nose is above the lip and a face has two eyes.
As the contents of these layers become increasingly abstract, these levels are also referred to as hidden layers. Through the interaction of several of these layers, “new” information can be formed between the layers, representing a kind of abstract representation of the original information or input signals. Even developers are therefore not, or only to a very limited extent, able to comprehend what the networks learn in the process or how they arrived at a specific result. One speaks here of the black box character of AI systems.
The advances in neural networks today are contributing significantly to the great leaps in development that AI systems have to offer today. However, the basic technology of neuronal networks is not new in itself: already in the early 1940s, the American researchers Warren McCulloch and Walter Pitts laid the foundation of neuro-informatics with a description of the first neuron model.
Due to their high performance in deep learning processes, neural networks are applied especially in those areas and problems in which extremely complex patterns are to be analysed in huge amounts of data. Neural networks are typically used in facial, object or speech recognition (natural language processing) and automatic text generation.
In the field of language processing, the implementation of neural networks leads to significant progress not only in the machine translation of a text into other languages, but also in the area of End2End text generation (natural language generation).
Great hopes are attached to the implementation of neural networks: End2End aims to enable an NLG system to acquire the process of text generation from the structured data to the finished text itself without human intervention.