I try to make it relatively short…
AI is first of all a generic term for pretty much everything that controls the behavior of a machine.Whether these are rules, programming, any sophisticated learning algorithms, does not matter at first. So most computer-controlled things can be called AI.
Machine learning includes a wide range of mechanisms that enable a machine to work statistically, rather than rules-based.Many different methods and algorithms are used, but the basic idea is ultimately to train the machine by means of examples instead of modeling the machine by hand.
In machine learning, as mentioned, various methods are used.Linear regression – Wikipedia is probably the “easiest”, best known method.In principle, a function is determined from the examples to predict the output at a given input. Super suitable for linear processes. Of this genus (“easy to understand” statistical methods) there are still some such as decision trees, K-Means, Naive Bayes and many more.
Neural networks, on the other hand, work slightly differently:
The idea behind neural networks is to create a kind of self-optimizing system.
You have as many hidden layers between input and output as you like. The planes also consist of nodes (neurons) that have edges in one direction (from input to output) to next-level nodes. Simplified, the training works like this: each edge gets random weights at the beginning. During the training, the data passes randomly (depending on the weight) through the nodes and ends up in the output. The paths between input and the “correct” output are gradually optimized by getting the edges that connect them more and more weights. This way, the system slowly optimizes itself. Seen in this way, it’s kind of a reinforcement learning…
Reinforcement learning (closely related to active learning) is applied in deep learning and is the concept that is also behind human learning processes.It is, by and large, a matter of taking random actions (exploration) and gradually rewarding the “right” actions (for which a special reward function is set up), so that the algorithm/AI is more likely to Running. Ergo: the successful actions are weighted more and thus executed more often. In principle, any learning method that uses “feedback” to further optimize itself is a learning process.
And last but not least, deep learning… Deep learning is a process in which the machine learns independently from the data with a lot of computing power in a kind of “smart brute force” by trying out a lot of.While classic training methods always require a ready-made input, coupled with the “correct” output – for example, 1000 images of each object that is to be recognized – i.e. “labels” or a “gold standard” gets the deep learning algorithm only a reward or cost function that tells it whether or not how good its output is. Over time, the machine should learn to interpret the input independently and deliver the right results. Sounds crazy, but in the end it’s a combination of all the above methods, coupled with very, very much computing power and good optimization. This is usually implemented with several levels of neural networks running as a chain one after the other.