How Unsupervised Machine Learning works

from | 22 May 2018 | Basics

Unsupervised Machine Learning helps - There are certain tasks that present enormous challenges for programmers. For example, if a programme is to be written that controls the optimal sequence of movements for robots, there are theoretically countless possibilities available. But which one brings the optimal solution?

With conventional programming languages like Python, it is almost impossible to programme tasks as complex as autonomously moving robots or cars. This is an ideal task for Unsupervised Machine Learning - sometimes also called Unsupervised Learning or UL for short. UL is, next to the Supervised Machine Learning one of the currently most important methods in the field of artificial intelligence.

In our article on "What is Artificial Intelligence"we explain the principles and individual methods in more detail.

What is Unsupervised Machine Learning?

Unsupervised machine learning can first be defined in distinction to supervised machine learning. While in supervised learning it is clear at the beginning what the result should look like - for example, whether an email message is labelled as spam or non-spam - the result of unsupervised machine learning is not known. Unsupervised Machine Learning not fixed from the beginning.

An illustrative example is provided by the following video, which shows how an AI tries to teach itself what "walking" is. It does this without having received any information in advance about how humans actually walk on two legs:

YouTube

By loading the video you accept YouTube's privacy policy.
Learn more

load Video

For the sake of simplicity, the difference between supervised and unsupervised learning can also be thought of as a Learning process with and without a teacher imagine. But is it even possible to learn something useful if there is no 'teacher'? The question can be answered with an unequivocal 'yes', at least with regard to machine learning.

An artificial Neural network may, for example, "play" with a given set of inputs to find possible Connections, patterns or Similarities in a given set of unlabelled data.

So how does the learning process take place if there are no training examples? The artificial neural networks are given a system to distinguish 'punishment' and 'reward'. If you are to learn the rules of a game, losing a game would be equivalent to punishment and winning a game would be equivalent to reward.

During the learning process, the artificial neural networks adjust the weights of the individual nodes in the network so that punishments are avoided in the future. Gradually, they can thus develop strategies to learn both rules of the game and winning strategies. The Areas of application for this form of learning are very broad and span the areas:

  • Economy
  • Research
  • Games

One of the well-known successes of Unsupervised Machine Learning comes from the last field. Google's AlphaGo was able to teach itself how to play Go without having the rules of the game explained to it in advance.

Even though they are shown separately here in the graphic: Unsupervised Machine Learning is not a method that must be used exclusively. So-called "hidden layers" can be pre-trained with Unsupervised Learning and then connected to a classic supervised network.

Supervised Machine Learning Overview
© datasciencecentral.com

The foundation for Unsupervised Machine Learning

Artificial neural networks and Deep Learning Algorithms are currently the best-known variants of unsupervised machine learning methods and thus form the basis. They are based on imitating the functioning of human nerve cells.

Simplified representation of a human nerve cell
Simplified representation of a human nerve cell. The neuron processes the input in the cell nucleus in the form of a signal that is interpreted or passed on.
Artificial neural network
Analogous to the way the nerve cell works, an artificial neural network processes an input signal that is processed according to certain rules.

Even if the "atomic" basic principle of Artificial neural networks (KNN) is very simple, they are ideal for making very complex tasks to cope with. Each Hub in a KNN behaves according to a certain rule or can be trained to do so.

The illustrated example shows a KNN with two so-called Hidden Layerswhich are located between the input and output layers. The term Deep Learning comes from the fact that these hidden layers have a very large "Depth", in other words: their number can become very high.

Unsupervised learning using the example of clustering

A subcategory of unsupervised machine learning is so-called "clustering", sometimes also called "clustering methods". At Clustering the aim is to group data without certain attributes according to certain criteria. This is relevant in marketing, for example, when it comes to segmenting customers.

Also in the area of Text mining and Data mining the clustering method is one of the most frequently used methods. From a large amount of text data, data sets with similar content can be grouped together - for example, all delivery notes for a certain product group or in genetic research when it comes to correlations of characteristics in genetic data. Clustering is also used in the areas of web mining, scientific questions, image processing and pattern recognition.

Semantic clustering
Examples of a semantic assignment according to a clustering procedure.
Semantic mapping in machine learning
Examples of a semantic assignment according to a clustering procedure (top) and different forms of clusters (bottom).

When it comes to the mathematical models that are applied in Unsupervised Learning, there are Four important categoriesinto which clustering procedures are usually divided:

  1. Partitioning procedures such as K-Means, K-Medoid, K-Mode or K-Median
  2. Probability-based methods for modelling clusters through probability distributions
  3. Density-based methods - Algorithms with data density and distance functions such as DBSCAN or SNN
  4. Hierarchical procedures - Hierarchical partitioning with agglomerative and divisive algorithms

Partitioning methods using the example of K-Means

A very popular method among the Unsupervised Machine Learning methods are Partitioning procedures. Such a procedure is a heuristic approach. This means that an exact result is not sought that can only be evaluated as "right" or "wrong". Rather, it is about a Process or a Functionin which the best Result approximate is achieved.

The goal is to sort the existing data into "k" different clusters. In the first step, a Random number of clusters to see what a possible result looks like.

Comparison of good versus bad clusters
The examples show how to distinguish good clusters from bad ones by measuring the distance between the individual points and adding them up.

The K-Means method is based on a fixed sequence in two steps that are repeated until a stable state is achieved:

K-Means Method in Machine Learning
K-Means Method in Machine Learning

Probability-based Procedure also have the goal of clustering data points, but assume the probability with which a certain object belongs to a certain cluster. This method is also characterised by a iterative Approach where each attempt is an approximation of the best solution.

Probability-based methods in machine learning
Basic idea of modelling clusters through probability distributions.

Density-based methods

Since clusters can be distinguished not only by their shape and size, but also by their density, density-based methods can lead to better results in some cases.

The decisive advantage with the Density-based clustering is that so-called "noise" can be filtered out. For example, there may be data objects that do not belong to any particular cluster. These can be filtered out within the framework of the density-based method.

Density-based methods in machine learning
Significant clusters can be formed via the data density, which have unusual shapes or which can filter out a certain "noise".

Hierarchical procedures

Last but not least, there are cases where Hierarchical procedures lead to the target. This method is always used when there are no global categories according to which data can be clustered. However, data can still be clustered according to certain parameters that follow hierarchical principles and can be grouped according to Size or Density can be staggered:

Hierarchical methods in machine learning

Unsupervised vs. semi-supervised learning vs. supervised learning

The distinction made at the beginning between supervised machine learning and unsupervised machine learning serves the purpose of better understanding, but is not sufficient in practice. To simplify, it can be said that supervised machine learning usually involves the prediction of a certain label such as "sick" / "healthy", "spam" / "non-spam" or "defective" / "functioning".

Often the Problems in practice but more complex and the solutions fall straight between the two methods. Depending on the application, it can therefore make sense to combine both methods. In this context, another term has become established: the method of "semi-supervised" learning.

At Semi-Supervised Learning only a few training data can have a label, while the majority is unlabelled. Depending on Data quality it can also happen that sometimes the data or the corresponding labels are incomplete. The first results in the Data analysis are then "estimated" or "well guessed". Only when the gaps have been filled in this way can the algorithm be improved on the basis of these results.

Concrete areas of application of Unsupervised Machine Learning

As mentioned earlier, Unsupervised Machine Learning first gained notoriety as part of the achievements of AlphaGo. Learning the rules of games as complex as Go or poker independently is considered a test for similarly complex Scenarios in the economy as in stock exchange trading, for example.

Even though many of the machine learning methods presented here are very abstract, in practice they fulfil very concrete goals. Be it in fraud detection (Fraud detection), the detection and prevention of customer churn (Churn Prediction Model) or the optimal planning of maintenance dates for machines and vehicles (predictive maintenance or preventive maintenance).

Even with complex image analyses such as the recognition of emotions (Affective Computing) or video analyses, Unsupervised Machine Learning can show its strengths, as the questions have an explorative character.

The big challenge is to choose the right method for the right application. This includes on the one hand Experience and on the other hand the corresponding Expertise. That's why in our Unsupervised Machine Learning training we provide detailed expertise that plays a crucial role in identifying the relevant machine learning methods for a specific use case.

The future potential of Unsupervised Machine Learning

Even if unsupervised machine learning is not as widespread as supervised machine learning, the method is no less important. Especially because machine learning in general is becoming one of the decisive factors for innovation and growth, methods such as supervised and unsupervised machine learning are also gaining in importance.

One difficulty in using UL is the assessment of the results. The analysis results are not simply "sick" or "healthy". Precisely because the output values are not fixed from the beginning, however, Unsupervised Machine Learning is ideal for Explorations and data exploration.

Especially with regard to the development of AI, Unsupervised Machine Learning is even considered a great beacon of hope. Above all, the ability that Algorithms through the use of Unsupervised Machine Learning independent Rules and Sample This makes it an extremely interesting tool in the development of AI and in its application in practice.

Author:inside

Michaela Tiedemann

Michaela Tiedemann has been part of the Alexander Thamm GmbH team since the early start-up days. She has actively shaped the development from a fast-moving, spontaneous start-up to a successful company. With the founding of her own family, a whole new chapter began for Michaela Tiedemann at the same time. Hanging up her job, however, was out of the question for the new mother. Instead, she developed a strategy to reconcile her job as Chief Marketing Officer with her role as a mother.

0 Comments

Submit a Comment

Your email address will not be published. Required fields are marked *