Machine Learning is a technology from the field of artificial intelligence. Find out what is behind Machine Learning and how it is used in companies.
Machine Learning is a technology from the field of artificial intelligence. Find out what is behind Machine Learning and how it is used in companies.
What is Machine Learning and how does it work?
The term machine learning describes mathematical methods with which machines can independently generate knowledge from empirical values. In simple terms, one could also say that with machine learning, computers can perform actions without having to be programmed for them beforehand. This approach can be assigned to the umbrella term "artificial intelligence".
From a technical point of view, machine learning is nothing more than man-made programs. The special feature, however, is that the algorithms used are able to recognise patterns in data independently. In order for machine learning programs to work, they must first be supplied with extensive data sets to "train" them. In this initial phase, the results are repeatedly checked by the developer. If necessary, the machine learning model is optimized and adjusted until it reaches the desired accuracy. The algorithm thus learns from dataset to dataset until it can fulfill its task without human intervention.
Incidentally, machine learning is not a new approach, but has been the subject of research for over 50 years. However, the methods only became practicable with increased (and affordable) computing capacities. At the same time, the amount of available data has increased exponentially. This created an excellent basis for the economically and strategically sensible use of machine learning.
What types of machine learning are there?
Machine Learning can be divided into three model groups based on the learning method used:
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning
Let's take a closer look at these categories of machine learning below.
In supervised learning, known data with an already existing logic is necessary. Two things are generated from this data:
- a training data set
- a test data set
The training data set enables the algorithm to learn the desired logic. It is then able to classify data with a certain similarity to the training set with the learned logic. The test data set, on the other hand, is required to check the performance of the algorithm. If appropriate tests provide results that can be considered good or accurate enough, the training process can be terminated. Machine Learning" was successful.
An application example for supervised learning is the recognition of objects on photos. For example, the goal of the machine learning process could be to automatically recognize dogs in pictures. In this case the training data set would consist of hundreds or thousands of photos, some of which actually show a dog. Of course, it is necessary to identify in advance which of the pictures this applies to. The test data set can then be used to evaluate the accuracy with which the object recognition is already taking place. If necessary, the algorithm can be refined until the desired accuracy is achieved.
Unsupervised learning is suitable for scenarios in which no known and logically structured training data for machine learning is available. Corresponding algorithms are able to independently recognize characteristic patterns in data and to divide them into groups. We are also talking about so-called clustering procedures. It is conceivable, for example, that the algorithm is supplied with all customer data and, in a second step, independently divides them into customer groups. The number of desired segments can either be specified or left entirely to the algorithm. In this case, machine learning is thus less directed in a certain direction.
Unsupervised learning, however, also requires human intervention. This is simply because the algorithm does not provide a reason why it has formed the clusters. The result must therefore be interpreted from a technical point of view and its usefulness critically questioned.
Another use case for unsupervised learning is dimensional reduction (also called principal component analysis). This machine learning approach serves to identify so-called features in existing data. In this context, features are nothing more than characteristics in which the data differ. As a simplified example, the descriptions of textiles can be used here. An extractable feature in this case would be the colour of the garments. Here, too, it is necessary to ask why the algorithm has used a particular feature as a distinguishing feature.
Encouraging learning is based on a form of human learning and is still at the beginning of its development in the field of machine learning. Through trial-and-error or "reward" and "punishment", a software agent learns how to act optimally in certain situations. In contrast to the machine learning types mentioned above, no training data is required in advance. Instead, this data is generated in numerous runs within a training environment. The goal of the process is to maximize the number of rewards in the Machine Learning simulation environment. However, the agent does not know in advance which action is best in which situation. Ultimately he develops a strategy on his own.
AI researchers see enormous potential in reinforcement learning, as machines could achieve human-like abilities to perform any intellectual task. In this case of machine learning we are also talking about "Artificial General Intelligence".
What is the difference between Machine Learning, Artificial Intelligence and Deep Learning?
In the context of machine learning a number of similar and related terms are used again and again. However, it is important for the basic understanding to be able to distinguish between them.
Let us start with Artificial Intelligence (AI). This term is nothing other than the English translation of "artificial intelligence" (AI). At the same time, AI is the collective term for all technologies used to provide intelligence services. One of these technologies is again Machine Learning. One level deeper is the so-called Deep Learning - a special form of machine learning. Since deep learning is an advanced and particularly relevant approach to artificial intelligence, we will now look at this area separately.
What is Deep Learning?
As part of Machine Learning, the advanced Deep Learning method is based on learning processes in the human brain. The high efficiency of this approach is achieved by using so-called artificial neural networks in combination with large amounts of data.
Deep learning algorithms are able to continuously link what has been learned with new information. They therefore learn continuously. In the course of time, the machine is then able to make independent decisions and question their effects. If no satisfactory result is achieved, the algorithm makes a new attempt. As a rule, humans do not intervene in this process. These are therefore "self-learning" systems. This is also the central difference to the standard methods of machine learning outlined above.
Machine Learning: Algorithms
Algorithms are a discipline of mathematics and computer science, without which machine learning cannot function. In simplified terms, one can imagine an algorithm as a recipe that describes exactly which steps have to be performed in a defined sequence. Applied to software development, these steps are commands of a programming language.
Some actions can be formulated very simply as an algorithm - for example, counting data records. However, this is more challenging for tasks such as handwriting recognition. In cases of this kind, it is simply not possible to map all eventualities manually in a program code. This is where machine learning methods come into play. We are also talking about "learning algorithms" here.
In contrast to classical algorithms that perform a defined task, machine learning algorithms have much more freedom. From a technical point of view, this is expressed by hundreds, thousands or even hundreds of thousands of so-called parameters. The learning process consists of adjusting these parameters in order to obtain results that are as correct and accurate as possible.
How are Big Data and Machine Learning related?
Big Data stands for very large amounts of data, which are also extremely complex, fast-moving and unstructured. In the context of machine learning, Big Data is of great importance in two respects. On the one hand, the data is required to train algorithms. On the other hand, Big Data cannot be evaluated using conventional methods of data processing, which is why advanced analyses based on artificial intelligence and machine learning must be used.
Today, the availability of data no longer poses a challenge. Unbelievably large amounts of information have long been generated every day - and the trend is rising. For the training of algorithms for machine learning, however, it is crucial to identify data with an appropriate quality. Incomplete or inaccurate data mean that machine learning models cannot deliver satisfactory results. Even more fatal is the fact that algorithms naturally apply incorrectly learned data incorrectly. The performance of machine learning systems therefore depends on the underlying data quality.
Which industries already use machine learning technologies?
Numerous companies are already using artificial intelligence productively in this country. The sub-discipline of machine learning in particular is now considered a significant factor for future profitability and competitiveness. According to Crisp Research, the approach is most widespread in the following industries:
- Consumer Goods
According to the study, chemicals, transport and logistics are also in the starting blocks and are expected to catch up shortly. Let's take a look at some use cases to illustrate the possible applications of machine learning in practice.
Use of Machine Learning in the industry
There are several concrete applications for machine learning in industry. Here, for example, the technology is used to evaluate data from production. This data is mainly recorded by sensors. Based on the knowledge gained (keyword: pattern recognition), quality improvements and more flexible production processes can be achieved.
Furthermore, machine learning is used for predictive maintenance - predictive and preventive maintenance. Based on sensor data, machine learning applications are able to detect impending machine failures in real time on the basis of specific status patterns. Consequently, maintenance can be requested in good time.
Use of machine learning in the service sector
In the service sector, machine learning has the potential to relieve employees of standardizable tasks. Chat offers, for example, which automatically answer simple customer inquiries or forward more difficult cases to the responsible employee, are well developed.
Automation of all inbound channels is also possible. Algorithms can now be applied not only to incoming letters, but also to social media posts and e-mails. Even handwritten documents can be included. In addition to customer master data, the artificial intelligence recognizes the purpose of the letter in particular. It can therefore classify the document within a few seconds and assign it to the right contact person.
How can machine learning be used economically?
Of course, the introduction of machine learning technologies must be economical. However, as this is a completely new field of technology for many companies, the benefits often cannot be evaluated concretely. It is therefore advisable to start in small steps. A good starting point is a manageable pilot project in the field of machine learning. At least the following questions should be clarified before starting:
- Which area of application is particularly suitable for the company?
- How can the necessary machine learning know-how be built up in the company?
- Which external support is necessary?
- Which machine learning models or model groups are relevant for the project?
- Is there a sufficient amount and quality of data available?
- Who bears responsibility if decisions are delegated to the machine?
All these aspects must be examined comprehensively and with expertise. In particular, the relevance, investment requirements and risks must be professionally assessed. Despite many questions, however, companies should not hesitate. It is definitely important to find a timely entry into the subject area of machine learning.
What does Machine Learning have to do with SAP?
For ERP providers such as SAP, artificial intelligence and especially machine learning methods are highly relevant. After all, ERP systems are considered the core of intelligent, automated processes. But the Walldorf-based software group SAP also relies heavily on artificial intelligence in the analytics area.
When looking at the current SAP portfolio, it quickly becomes clear that numerous AI solution modules are already available. In customer support, for example, "SAP Conversational AI", a bot platform that understands natural language, is used. Another application area is "SAP Intelligent Robotic Process Automation". Here, metadata-based bots ensure that standard processes are run through automatically.
Other examples of machine learning from the SAP world are:
- SAP Cash Application: automatic reconciliation of open receivables and incoming payments
- SAP Predictive Analytics: Making predictions about future business results
- SAP Service Ticket Intelligence: automatic classification and forwarding of customer tickets including determination of possible solutions and answers
- SAP Customer Retention: Forecasting Customer Behavior
What is the SAP Leonardo Machine Learning Foundation?
The SAP Leonardo Machine Learning Foundation plays an important role in the machine learning area at SAP. This is a platform with which self-learning applications can be developed, operated and used. No special IT knowledge is required to create the algorithms. SAP Leonardo Machine Learning Foundation is cloud-based, can be seamlessly integrated into SAP systems, and contains several Web services. Possible usage scenarios include:
- Image classification (e.g. identification of a vehicle category based on photos)
- Object recognition on photos (e.g. focusing on specific image content, for example in the validation of ID documents)
- Text recognition within images (e.g. for reading material numbers on photos or scanning labels)
- Text classification (e.g. identification of positive and critical moods in social media texts)
- Recognition of the national language of a text with translation
It is advantageous that companies can fall back on already existing models in the area of machine learning if required. These then only need to be adapted using their own data. This is a particularly recommendable way to get started in machine learning.
Meinolf Schäfer, Senior Director Sales & MarketingDo you have any questions? I'd be happy to help.
+49 2241 8845-623