For machines to learn, you first have to provide them with a suitable basis of relevant knowledge. While devices in deep learning know to learn independently, in traditional machine learning, the first imprint is taken over by humans. Today, machine learning is increasingly deep learning.
To learn, a basis of data sets as a kind of knowledge database is necessary, which provides central information for pattern recognition. Artificial Intelligence can either access data from an online database such as Wikipedia or supplied with its database set up and used offline. While the data sets that artificial intelligence can find and evaluate online can hardly be created offline, the data can be more strictly controlled offline. Machine learning takes on recognizing individual data strands from the data sets and evaluating them. At the same time, it summarizes the data strands as groups and clusters.
Incidentally, before actual use with available data sets, an AI is always “fed” with specific training and sample data. The behavior of your algorithms compared to these training data sets is carefully checked and evaluated to determine whether it is qualitatively appropriate and precise. This process is called model training. The model training can be viewed as a simulation that precedes the “real world” use with unknown data.
Algorithms determine how and to what extent patterns can be recognized and evaluated. You define the steps required to evaluate a task. Machine learning always requires a central machine learning algorithm and, in addition to it, various other algorithms, which in turn contain a series of predefined actions. Different possible solutions depend on a specific task, its complexity, and the type of problem. The central algorithms are the
- Decision Tree Algorithm: This algorithm is a kind of tree diagram in which various decisions are recorded according to which the AI reacts. Banks and companies active in the financial sector, for example, use decision trees to determine whether an investment is worthwhile.
- Random Forest Algorithm: A randomized Classification and Regression Tree (CART) is created for every conceivable scenario of a concrete situation. This means that an accurate prediction can be made for every reasonable possibility of how a specific problem could turn out. In simplified terms, this algorithm can be imagined as a thought construct designed as part of a chess game. All options are played through mentally before an actual one is made.
- K-Means Algorithm: This algorithm solves clustering problems, i.e., the grouping of data strings. They classify data and subdivide it accordingly. For example, K-Means can cluster visitors to a website into actual people and bots.
Algorithms can either make outputs simultaneously as inputs ( batch learning ) or learn sequentially, i.e., evaluate the input data first and make outputs staggered.
So-called learning categories distinguish how algorithms analyze and evaluate data. They are models according to which algorithms develop and which focus they have. Machine learning can
- unsupervised learning ( supervised ),
- supervised learning ( unsupervised ),
- semi-supervised learning or
- reinforcing learning ( reinforcement ).
The difference between supervised and unsupervised learning is that data is manually assigned to appropriate model groups of the algorithms with supervision. If there is no monitoring, the machines automatically compare data with patterns and form independent model groups. Supervised learning lends itself to measurable predictions such as risk assessment, accurate/false estimation, or spell checking. Unsupervised learning is mainly used when evaluating larger data sets. Partially supervised learning describes the partly manual and partly automated formation of model groups. An example area of application is face and object recognition, evaluation, and adaptation.
In reinforcement or reinforcement learning, algorithms learn what data processing is desired through “rewards” and “punishments.” This category of education bears a certain parallel to human knowledge. Reinforced learning is mainly used in autonomous driving, autonomous robotics, or in games (i.e., for fun AIs). Research also hopes to become a general, i.e., superordinate, artificial intelligence that can become completely independent. Reinforced learning can be implemented as Temporal Difference Learning (TD-Learning), in which rewards are given directly in response to appropriate behavior; or according to the Monte Carlo principle, after which the prize is only awarded at the very end of a mission.