Deep Learning is a branch of AI. Artificial neural networks are used to autonomously learn from existing data.
How does a computer recognize individual faces, how can it talk to a person? How do autonomous cars drive, avoid obstacles in a fraction of a second or recognize traffic signs? Artificial intelligence is concerned with the solution of such questions - and here especially the subfield of deep learning.
Deep learning is based on "deep" artificial neural networks (DNN). The advantage over conventional software solutions is that these artificial neural networks are capable of learning and can process huge amounts of data. And the larger the amount of data, the more and better the systems can learn.
Deep Learning takes as its model the structure and type of information processing of the biological, human brain - the nerve cells and their connections with each other. However, Deep Learning is not about the exact reproduction of biological networks. Deep learning experts are developing models that can process information in a similar way to a human being - it is about machine intelligence.
With the help of artificial neural networks, computers solve even the most complex tasks for which no conventional computer programs can be written.
For simple problems, mathematical rules can be formulated that computers can easily process. It is far more difficult to solve problems with the help of computers that cannot be described with mathematical rules. These include tasks that people can solve intuitively - such as speech or face recognition.
The technology of deep learning has made enormous progress in recent years. This is mainly due to the increased computing power of computers.
Deep Learning is a sub-area of Machine Learning and thus of AI. The technology uses artificial neural networks. With the help of these networks it can process enormous amounts of data. The technology learns from this data. In this way, it independently solves complex questions and problems.
Artificial neural networks are the basis of the deep learning process. They consist of artificial neurons and their connections.
An artificial neuron is a simple arithmetic unit, also called a unit or node. The connections between the individual neurons are abstract structures, also known as edges. Data is transferred via these connections.
The connections are weighted. They are therefore of different strengths. The specific weight inhibits or excites the transmission of data.
The individual neurons are arranged in layers. Each neural network has an input layer and an output layer. In between there is one, but usually several hidden layers or intermediate layers.
There can be several hundred such processing layers. The more layers there are, the more computing power is required for deep learning. The number of neurons and layers depends on the task and can be precisely defined.
The data is passed from one layer to the next, starting with the input layer. It is the starting point of information processing. The data are processed in the neurons.
One neuron of the input layer passes its calculations to all neurons of the next layer. If the information is passed on exclusively in the direction of the output layer, this is called a feed-forward network.
However, there are also networks in which the data is passed on in other directions - for example "Feed-back networks" or "Fully connected networks"
Their structure makes "deep neural networks" very powerful. But how do such systems learn? What exactly happens in "deep learning"?
The learning process in machine learning is still quite easy to understand. The programmers provide their algorithm with certain input and output values during training. All values are known and correct (supervised learning).
After some calculation processes the algorithm can recognize relations and patterns between the input and output values. Thus it learns to apply these associations to new inputs.
The IT experts can then compare the results with the known and correct output values. This makes it easy for developers to check how good the results are. The results in turn are the basis for the next round of training with the algorithm.
With the deep learning approach, on the other hand, the path to the result can no longer be traced from the outside - the systems resemble a "black box". The learning processes take place in hidden processing layers of the artificial neural network (unsupervised learning).
A combination of different methods is usually used for training (supervised and unsupervised learning). Sometimes such a neural network is provided with input and output data for deep learning, but sometimes only input data and a task.
In such a case, the network tries to autonomously establish relationships between the information, recognize patterns and carry out segmentation (clustering).
Such an artificial neural network learns by changing and re-weighting the connection weights between the neurons. Any errors that occur in the evaluation of information are detected, taken into account in the next training rounds - and performance is continuously improved.
The deep learning programm AlphaZero, for example, shows what this means. It was introduced in 2017 and plays the board game Go and chess, among others.
Unlike its predecessors, however, the developers did not train the program with an extensive game database. AlphaZero received the rules of the games for training purposes alone - and then learned by playing against itself.
The result: The AI developed its own game strategies and even found new and surprising ways to solve problems that had never been found by humans before.
Deep Learning is an advanced subfield of Machine Learning.
While machine learning is about pattern recognition with the help of special algorithms and not neural networks, deep learning uses artificial neural networks.
Deep learning can solve even more complex problems than machine learning, especially with regard to unstructured data and very large amounts of data.
When large amounts of data (big data) and a lot of computing power (high performance CPUs and GPUs) are available, deep learning technologies can show their strengths.
Today, the technology is mainly used in image recognition as well as text and speech recognition. The processing of images plays a role, for example, in Facebook's face recognition, Google's image search or Amazon's Alexa language assistant.
Large amounts of data also accumulate in factories. This applies, for example, when smart sensors monitor machines. With the help of AI, the best maintenance point can then be predicted - according to actual demand and before the machine breaks down due to a malfunction.
It is similar in road traffic, where cameras and sensors of autonomous or semi-autonomous vehicles collect countless data every second. Processing this amount of data and deriving actions in real-time - this is only possible with Deep Learning.
Some areas in which Deep Learning is used:
Automotive engineering (e.g. self-driving systems, driver assistance systems)
The SAP assistants of Nooxit use algorithms of Machine Learning and Deep Learning technologies. The reversal assistant, for example, finds potential reversals by previously training with large amounts of data. In this way, the AI learns when reversals are likely - and can warn users in real-time.