Machine Learning - three important learning methods at a glance
After giving a general insight into Machine Learning in a previous article, I would like to talk about a different topic. And specifically on which methods or sub-disciplines are actually under this big topic of AI. This does not refer to mathematical methods such as neural networks, deep learning or other statistical learning methods, but to the categorization of these methods.
Do I need to monitor my AI? Supervised and unsupervised learning
Basically, the statistical learning methods behind the term AI can be divided into two groups, the so-called supervised learning and unsupervised learning.
The term monitoring is somewhat misleading and refers to whether the learning progress of the neural network can be monitored. In concrete terms, this means that in supervised learning a relationship between input variables and (known!) output variables is to be found. These are provided to the algorithm together, for example in the form of dog and cat images, each titled as such, or images of acceptable and reject products that also bear the respective label.
Figure 2 shows an example of supervised learning, where the relationship between the shape of a geometric figure and the number of its vertices is to be learned.
In contrast, in unsupervised learning, such a target and knowledge about contexts is not fixed. Rather, the algorithm itself is intended to find relationships in data sets in order to learn more about the structure of a data set. And consequently to gain information that cannot be found by pure human analysis, or only with great difficulty.
The two most important applications are clustering, i.e. grouping, or anomaly detection. Figure 3 shows an example of the application of a cluster algorithm. A group of geometric objects is given, which are to be divided into two groups so that the objects within the groups are as similar as possible but as different as possible from those of the other group. The possible group assignments are given by the two colors on the right.
A common example of anomaly detection is found in online shopping called fraud detection, which is the detection of conspicuous shoppers or purchases that indicate abuse. In addition, cluster detection is often used in medicine to find and group subjects with similar medical conditions.
Machines learn like children: reinforcement learning
The very simple examples so far show the basic features of machine learning, but do not yet seem really "intelligent". Therefore, reinforcement learning is used in many applications. This idea is based on how humans or animals learn. For this, the problem must be formulated in such a way that there is a so-called agent that is controlled by the AI, is subject to certain rules, and has certain possibilities for action. In addition, it must be possible to evaluate whether the task was solved well or poorly after each play-through of a scenario. Basically, there are two possible behaviors for agents: One is the exploration of new knowledge (exploration) and the other is the exploitation of existing knowledge (exploitation).
Training occurs in multiple phases consisting of many individual runs of one or more scenarios. In the early stages, the AI first tries to explore the available options through various operations and find references for "good" and "bad" actions through a large number of runs, based on the evaluation. After a phase is completed, the "best" runs are used as the basis for regular training of the AI. This is where supervised learning can be found again. In the later stages, the AI will try to exploit its knowledge more and more and will prefer more and more often a behavior that will surely lead to a better result.
One of the most impressive recent examples of reinforcement learning comes from the OpenAI team. These researchers have managed to teach a robotic hand to solve a Rubik's cube without any specific programming of the hand or how the Rubik's cube works. In the beginning, the AI "randomly" controlled the individual motors of the hand. The result was evaluated according to whether the Rubik's Cube was closer to a solution than at the beginning of the run. You can find the video here!
Another example of reinforcement learning can be found in the simulation of a hide-and-seek game, where two AI teams compete against each other and find ways to exploit bugs in the game engine itself to win the game while solving the problems involved.
Artificial Intelligence vs General Artificial Intelligence
As impressive as the aforementioned examples of AI use are, they all share the same shortcoming. Each of these examples required teams of experts who spent months working on the problems and had access to very large computing resources.
Therefore, one of the biggest goals of leading AI researchers, for example around the Deepmind team, is to create a so-called General Artificial Intelligence. That is, an AI system that, similar to humans, can understand a problem on its own and, so to speak, learn itself autodidactically.
Until such general artificial intelligences exist and are widely available, AI is a tool that is very powerful, but can also be very fragile and always involves large initial efforts for problem definition, setup, training, and validation.
Here, we as pragmatic industries together with the experts from pragmatic minds support our customers in the selection and implementation of AI projects. We specialize in machine data, i.e. product or production data that is generated on the store floor.