A definition that has existed since the term artificial intelligence was introduced in the 1950s is as follows: artificial intelligence is the ability of a machine to imitate intelligent human behavior. However, the machine only becomes interesting for truly efficient use when it can match or even surpass human performance. Because then we humans will have the opportunity to hand over repetitive tasks or have things done faster, safer or more error-free by computers.

Engineers and computer scientists currently speak of artificial intelligence when they can train a computer in such a way that it learns and imitates certain human abilities - for example, the recognition of patterns and objects or the interpretation of language. This expands the above definition:

"Artificial intelligence is the ability of a technical system to emulate or surpass intelligent human behavior - by training a computer to learn the desired behavior."

A technical system should therefore be enabled to perform a specific task. Traditionally, this has meant creating a program for a computer that processes data and delivers the desired output. This is a deterministic process.

Current AI technologies - machine learning, deep learning and reinforcement learning - reverse this approach: you feed a computer with data and train it using mathematical procedures - which vary depending on the method and model - to determine a meaningful result from this data. Computers generate the desired output using such methods by learning the "correct" model parameters, but the statement that a computer programs itself is not true.

The process of processing and interpreting the data does not consist of a specific sequence of operations like a traditional program or algorithm. Ultimately, AI models and procedures are essentially "black boxes" that learn and are trained to process and interpret data in a specific way using mathematical methods. This is not a hard-deterministic process, and a residual error remains, just as in human recognition and processing procedures. This residual error must be minimized and the process speed increased so that an AI system can match or surpass human capabilities for certain selected tasks.

Machine learning - the technology for data analysis

Machine learning is a data analysis technique that gives computers the ability to learn from experience. Machine learning algorithms use computational methods to "learn" information directly from data without using a given equation as a model. As the number of samples available for learning increases, these algorithms adaptively increase their performance.

Machine learning algorithms find natural patterns in data that can generate insights and support decision-making and forecasting. Areas of application include medical diagnoses, stock market trading and forecasting electricity loads.

Machine learning uses two types of techniques:

• supervised learning: creating and training models for predictions using known input and response data;
• unsupervised learning, which is used to find hidden patterns or internal structures in input data.

A supervised machine learning algorithm uses a known set of input and output data, called the training data set, to train a model so that it can make informed predictions of response values for new data sets.
Using larger training datasets and optimizing model hyperparameters can often increase the predictive power of the model and ensure that it generalizes well to new datasets.

Classification and regression techniques

In supervised learning, classification and regression techniques are used to develop the predictive models.

• Classification techniques predict discrete outputs - for example, whether an email is genuine or spam, or whether a tumor is malignant or benign. Classification models classify input data into categories. Typical applications include medical imaging, speech recognition and credit scoring. Classification is used when data can be tagged with keywords, categorized or divided into specific groups or classes.
• Regression techniques predict continuous outputs - for example, temperature changes or fluctuations in energy demand. Typical applications are the prediction of electricity loads and algorithmic trading. Regression techniques are used in the case of data ranges or spectra, or when the output is a real number, such as a temperature or the time until a device fails.

Unsupervised machine learning finds patterns or internal structures in data. It is used to draw conclusions from (possibly very large) amounts of data. The input data is not labeled and there is no classified output data with which to train a model.

Clustering is the most common unsupervised learning technique. It is used for exploratory data analysis to find hidden patterns or groupings in data. Applications for cluster analysis include gene sequence analysis, market research and object recognition.

Deep learning - a specialized form of machine learning

Deep learning is a specialized form of machine learning with which computers acquire the ability to learn from examples. In contrast to "classic" machine learning, where meaningful characteristics or features first have to be extracted from the data, in deep learning a computer model learns how to perform classification tasks directly from unstructured raw data such as images, text or acoustic data.

Deep learning methods use neural network architectures with many layers. This is why deep learning models are often referred to as deep neural networks. The term "deep" refers to the number of hidden layers of the neural network; deep networks contain up to 150 hidden layers.

Deep learning models require very large, labeled datasets for training, for example millions of images for image recognition, and learn to recognize features directly from the data without the need for manual feature extraction. The more training data sets are available, the higher the recognition accuracy. Deep learning also requires a great deal of computing power, which is often provided by high-performance GPUs.

Due to these requirements, it has only been used in practice for a short time, although the theory of deep learning was already established in the 1980s. The recognition accuracy of deep learning algorithms now exceeds that of humans in some fields of application, for example when classifying objects in images.

Examples of deep learning in practice

• Automated driving: Automatic detection of objects such as stop signs, traffic lights and pedestrians
• Aerospace: identification of objects, zones and areas by satellite
• Image analysis in medical research: automatic detection of cancer cells, interpretation of CTs and MRIs
• Industrial automation: increasing work safety on machines by recognizing people or objects in the work area
• Interpretation of speech and text (electronics): Automated listening and speaking, automated translations.

Reinforcement learning - AI for virtual models

Reinforcement learning is a form of machine learning in which a computer learns to perform a task through repeated interactions with a dynamic environment. In this way, control and decision-making systems can be implemented for complex systems such as robots and autonomous systems.

An iterative trial-and-error approach is used to generate data from a dynamic system, which the computer uses to learn to make decisions and find the best sequence of actions, for example for a sequence of movements or a sequence of game moves. This is done without human intervention and without the computer being explicitly programmed for this task. A well-known example of reinforcement learning is AlphaGo, the first computer program to beat a world champion in the game of Go.

The aim of a reinforcement learning algorithm is to find a strategy that leads to the optimum result. To do this, so-called agent software interacts with an environment in order to explore it and learn from it. The agent consists of a strategy that assigns an input state to an output action and a training algorithm that updates and optimizes this strategy. The agent measures its performance in relation to the task goals using a (positive or negative) reward signal calculated from the environment. The algorithm updates the strategy to maximize the long-term reward signal.

Examples of training algorithms are Deep Q-Nets, Actor-Critic and Deep Deterministic Policy Gradients. Strategies can be represented by deep neural networks, polynomials and lookup tables.

For applications such as robotics and autonomous systems, training an RL algorithm in the real world with real hardware can be expensive and dangerous. Therefore, reinforcement learning usually uses virtual models of the environment that generate data through simulations.

The article was provided by MathWorks.