Overview

Machine learning processes can be characterized by the type of feedback from the environment that is intended to improve the behaviour of the system (“feedback”). This distinguishes between the three basic types of machine learning:


more on that

In supervised learning, the training data consists of labeled examples – i.e., pairs of inputs with the corresponding target values for the output, such as labels or function values.

The model is adjusted to minimize the deviation between the actual and target values – the so-called loss. The aim is to gradually bring the future output closer to the target and reduce the error in further predictions.

As a rule, the trained model is tested after the learning process before it is used in practice. For this evaluation, part of the available data – the so-called test data – that was not used for training is retained.


more on that

A prior classification or labeling of the data is not necessary here. The task is to find suitable classifications (“clustering”) or representations (“autoregression”), for example.

However, the role of the human being is still central; it consists, among other things, in the pre-selection of relevant characteristics and the way in which the “generalization” takes place: What does “similar” actually mean? What should be differentiated? Unsupervised learning methods can be used, for example, to determine typical feature characteristics or to recognize repetitions or exceptions. It is not a matter of assigning patterns to existing categories, but initially of producing class classifications, e.g. finding “clusters” in a data set.


more on that

The reinforcement learning method refers to systems that are designed to interact increasingly successfully in an environment (“adaptive agent programs”). They receive a reward for actions that they select from the available range of actions depending on the situation identified.

Reinforcement learning does not require a correct “target specification”. A simple numerical evaluation is sufficient for the system. In future interactions, the “reward” obtained should tend to be greater. The aim is to automatically develop a strategy that is as successful as possible (“policy”, sometimes also called ‘control’ or “tactics”).


Overview

  • Supervised learning: The training data is “labeled”. The difference between the given “correct” values and the values estimated by the system is minimized.

  • Unsupervised learning: The “costs” of a model (e.g. prediction errors, resource consumption) are minimized. This “error measure” is calculated internally, which means that the model is evaluated autonomously without additional external information.

  • Reinforcement learning: The system receives rewards/punishments when interacting with its environment. The ML system (an “agent”) maximizes future rewards.

Examples

  • “Supervised learning” - distinguishing between images with skin cancer or harmless pigmentation.
  • “Unsupervised learning” - optimizing the spatial design of a supermarket or its pricing policy, e.g. by recognizing which goods are often bought together. A special offer on red fruit jelly, for example, could be offset by a slightly more expensive vanilla sauce. Here, what the algorithm should pay attention to when grouping, e.g. “bought at the same time”, must be anchored in the ML system.
  • “Reinforcement learning” - Which advertising is particularly effective for which customer types, e.g. because it captures the customer's attention or leads to a purchase decision?

Modern systems combine various machine learning approaches, resulting in complex modular systems that combine different methods, e.g. in generative AI systems such as chatbots with large language models such as “Bing Copilot”, “ChatGPT”, ‘Claude’ and “Gemini”.