Machine Learning: An explanation of how it works, subcategories and possible use cases
Machine learning is a subcategory of artificial intelligence where recurring patterns within complex structures are recognized and made usable. In comparison to classic algorithms, which represent a calculation rule according to a fixed scheme, machine learning can independently gather experience and use it to make predictions for the future. Building on the feedback from the actions, the models are able to learn independently and make improved decisions based on this. This allows computers to master tasks without explicit and task-specific programming.
In this article we describe how machine learning fits into the world of artificial intelligence and how it works. We will also refer to the different subcategories of machine learning. To do this, we will get to know how supervised, unsupervised and reinforcement learning work in more detail, look at current use cases and work out the advantages and disadvantages.
Integration of machine learning in the field of artificial intelligence
Definition of artificial intelligence: There are many different definitions for the term “artificial intelligence”. One approach is to first define human intelligence in general and derive the term AI from this.
intelligence: The ability to create creative connections between already known knowledge and new impressions.
The AI should be able to replicate this and needs three characteristics to do this:
- Understanding of language & text
- Storage, processing and evaluation of recorded information
- Ability to learn & react flexibly to existing circumstances
Overall, AI is made up of the sub-areas of action, perception and learning, although we will focus on learning in this entry.
The category of learning compares the most recently collected data with the previous level of knowledge in order to improve one's own output. The program is designed to learn independently based on new experience. In the following we will focus on machine learning.
Definition and functionality of machine learning
Machine learning is a subcategory of artificial intelligence and enables data-based predictions to be made. For this purpose, experience is collected using past problems and the solutions derived from them and new solutions for the current problem are developed. The more data is available, the better the model will be at recognizing patterns and making targeted statements. The model does not have to work within predetermined rules, but can also independently find the optimal solution.
An important measure of the quality of the output is generalizability, which means that the model is able to generate the correct output even with unknown input data. Since the model is usually confronted with new data, it is important to apply the experienced patterns to the new data.
The graphic below shows how the subcategories of machine learning differ from the classic heuristic rules.
Heuristic rules: Human knowledge is used to create specific rules and instructions for solving problems based on assumptions and experience. A major disadvantage of this method is that there is no learning process and therefore cannot react to new data. The advantage is that this approach is less complex and easier to interpret due to greater transparency.
Machine learning: Enables technical systems to learn patterns and develop further. In contrast to heuristic rules, it is able to find suitable solutions even for unknown data. In general, it can be divided into three main categories:
-
- Unsupervised: recognizing connections
- Supervised: Removing connections with a clear goal
- Reinforcement: Independent recognition of correct connections
Unsupervised learning: explanation, areas of application, advantages and disadvantages
What is Unsupervised Learning?
Unsupervised leaning is a central method of machine learning in which the model independently recognizes structures in the data. In contrast to supervised learning, no pre-categorized data is available; instead, similarities are found within the input data. This is intended to identify hidden patterns and gain valuable insights that people would not have noticed.
Where do you use unsupervised learning?
Since unsupervised learning works with data that does not have target values, a large number of use cases open up for this technology. For this purpose, cluster analysis, dimension reduction and anomaly detection are presented below:
Clustering
Clustering is the process of organizing similar data into a group without providing precise information about each group. The individual elements within a group should be as similar as possible. For example, customers who have purchased similar products in the past can be put together into a group and then offered the same product.
Dimensional reduction
Data sets often have a large number of different features, which are referred to as dimensions. Because the different columns correlate with each other, it is possible to minimize the number of columns, leaving only two or three dimensions. These can then be visualized using simple graphs. This is used, for example, in blood cancer research, as the data often contains more than 30 variables. These are reduced to three dimensions and can then be visualized to recognize patterns.
Anomaly detection
Using unsupervised learning, it is possible to identify individual outliers in data sets that differ from the normal data set. An example here is the detection of credit card fraud by identifying transactions that are unusual for a customer.
Advantages
Lower expense: Training data does not have to be labeled
speed: Since the data does not have to be labeled and simpler models can often be used compared to supervised learning, the speed can be increased
Scalability: It is possible to apply these algorithms to a large amount of data
Disadvantages
Limitation: Only allows classification tasks
Evaluability: Since there are no clear output values, the possibility of assessability is limited
Human interpretation: There must be a human-led evaluation of the clusters created
Supervised learning: explanation, areas of application, advantages and disadvantages
What is Supervised Learning?
Supervised learning belongs to the category of machine learning and requires labeled data compared to unsupervised learning. This means that the input and output are labeled to measure and improve results. The algorithm recognizes patterns from training data sets and applies them to new inputs. Therefore, the quality of the training data is particularly relevant for this type of algorithm.
The previous graphic shows the basic principle of supervised learning. The model is first pre-trained with labeled training data (here, for example, the geometry with the right terms) and can then apply what it has learned to new data. According to the example in the graphic, the correct terms can be assigned to the geometric shapes from the new data set.
Where do you use supervised learning?
For a more precise understanding of supervised learning, the most common areas of application of this technology are presented below. In this context, classification and regression are also discussed.
classification
The model tries to assign the correct label to the input data in order to solve decision-making problems. To achieve this, a function is sought that divides the data into classes. This enables prediction of discrete values. This procedure can be used, for example, to classify spam emails.
Regression
The aim of regression is to analyze the relationship between variables to enable prediction. To do this, a mapping function must be found with the help of those the input can be mapped to the continuous output. This is set bFor example, for sales forecasts or property price predictions.
What are the advantages and disadvantages of supervised learning?
After we have explained the basic mechanism and areas of application of supervised learning, we will look at the advantages and disadvantages below.
Advantages
transparency: Unlike unsupervised learning, the result here is easier to understand
Security: Controlled learning environment by using labeled data
Disadvantages
Complex pre-processing: The data must be labeled by human intelligence
maintenance: Requires regular updates
Limited: In contrast to unsupervised learning, no unknown information can be obtained from the data set
Reinforcement learning: explanation, areas of application, advantages and disadvantages
What is Reinforcement Learning?
Reinforcement learning is the third category of machine learning and aims to ensure that an agent learns a strategy independently. He uses this to decide which action should be carried out in which situation in order to maximize the reward function. The basic functionality is shown in the graphic below:
At every time t the agent receives a representation of its environment, which is called State St referred to as. Depending on this state, the agent chooses one Action At out of. In the next period (t+1) the environment has changed and now has the State S(t+1). The agent receives another one Reward R(t+1) for the Action At out of State St. This process repeats itself; the agent learns from every interaction with the environment and adapts its strategy accordingly. Since Reinforcement Learning is a broad umbrella term, we will go into the individual subcategories (MDP & Bandits) in more detail in another entry.
Where do you use reinforcement learning?
Since reinforcement learning is a broad term and has many subcategories, we will focus on the most important areas of application below:
Autonomous driving: Reinforcement learning can be used to make cars drive autonomously through pre-training in a simulated environment. This allows the car to react adequately to all possible situations and obstacles on the road. For example, reinforcement learning is used to learn automatic parking strategies or overtaking strategies while avoiding collisions. Q-Learning (a type of RL) can also be used to optimize the lane changing process.
Trading: RL can be used to predict stock price. To do this, the agent learns a strategy such as: B. should trade stocks to maximize profits.
Advertising strategies: Here RL is used to suggest the optimal product for the customer. For this purpose, data about historical customer behavior as well as other customer data, such as: B. demographic information is used.
What are the advantages and disadvantages of reinforcement learning?
Advantages
adaptability: Can be used in environments where no labeled training data is available
performance: RL can solve complex tasks effectively because it can react to the respective environment
Disadvantages
Long learning process: Agent requires many interactions with the environment until a suitable policy is found
Low transparency: Since the decisions result from an interactive process, the resulting output is often difficult to understand
In this article we have given a general introduction to machine learning and explained the basic mechanisms. Building on this level of knowledge, we can now delve deeper into the individual models. For this purpose, selected algorithms for the three main categories (supervised, unsupervised, reinforcement learning) are explained in the following article.
If you would like to learn more about real-world applications of AI, please take a look at our AI products. For example, we use unsupervised learning algorithms in our product blue.LEAD to combine website content into clusters (click here). If you want to find out more about reinforcement learning, take a look at the use case for blue.ACTION (click here).
EN 
DE
A very informative summary of what ML is and what categories it can be divided into!