The common basis for the work at the ML2R is the research and development of modular applications of Machine Learning (ML). In modular Machine Learning, systems are built up from linked individual modules that can be used intuitively and reused flexibly. This makes ML procedures more flexible and increases their transparency and traceability.
The basic idea of modular ML is to divide a complex question into separate steps. To answer the individual questions, modules are used which the developer can at best pull like finished tools from a case and reuse at the next opportunity for a similar problem. In practice, however, even the decomposition into single steps is often a great challenge for human developers. Machine Learning methods can help to reliably build an ML architecture from a chain of individual modules.
Examples of ML tasks, which consist of many specialized subtasks and can be built from modules, are systems for automating cognitive processes, such as the automatic processing of requests to an insurance company. First, the system needs to process and understand documents, then categorize the requests, extract and summarize relevant information, make the right follow-up decisions and generate a response.
Modular architectures are already established in software development. The aim of the ML2R team is to adapt this procedure for ML applications and to explore also theoretically how individual ML modules can be automatically linked to analysis chains. Such a modular system for Machine Learning will be more comprehensible to humans, easier to adapt and reuse for different applications and integrate various forms of knowledge.
Machine Learning with restricted resources makes it possible to reliably perform calculations using Machine Learning even on small devices such as smartphones or directly in sensors. For a long time, computing resources could be ignored. Meanwhile, the models of Machine Learning are thought together with different computer architectures up to quantum computing.
In modern, automated production facilities, large amounts of data are generated in addition to the actual products. Operating data of machines and industrial robots or measurement results of sensors are sources of such data.
The evaluation of this data treasure trove holds enormous optimization potential: processes can be accelerated, quality can be improved, machines can be maintained preventatively and robots can be employed more efficiently.
Such applications often rely on Machine Learning methods for processing large amounts of data. Artificial neural networks in particular are able to learn from a large sets of sample data and to develop models. These models then help production engineers, for example, to optimize manufacturing. However, such applications require very powerful computers.
Especially for applications in industry 4.0 and logistics, it is not always possible – for example due to a lack of bandwidth for data transmission – to send all information to a powerful central computer. Instead, the devices that record the data, such as sensors, have to take over part of the processing and analysis.
The developers at the ML2R Competence Center are working on making Machine Learning also available on devices with limited computing power and memory. In addition to sensors and hardware that can be linked within the Internet of Things, such devices are smartphones or mobile computers. One research approach is to simplify the ML algorithms so that they require less memory and computing capacity. Another course of action focuses on the development of hardware and software that is optimized for specific learning tasks.
ML procedures function well if they can learn with a lot of data. The situation is different in situations where little or no data is available. In logistics, for example, weather influences, construction measures or accidents can cause exceptional situations for which no sufficiently large sets of data are available.
The example of logistics applications serves to illustrate another challenge of ML: Data and knowledge originate from different sources and have very diverse formats. Containers used for the transport of goods are increasingly equipped with sensors that send information on the current position as well as on pressure and temperature inside. This sensor data must then be linked with knowledge from other information sources, such as traffic planning systems and weather data. Sometimes, however, knowledge is only available in the form of experience or intuition of human experts.
The researchers of the ML2R Competence Center are working on methods that make heterogeneous sources of data and complex knowledge usable for ML. To this end, they examine methods for processing and presenting knowledge in a homogeneous form. In addition, data-based ML forms, such as deep learning, are to be linked with other forms of learning based on explicit and logical knowledge. For logistics applications, for example, the goal is to develop a system that makes the complex interrelationships in logistics planning transparent for people and at the same time uses the potential of ML to optimize processes.
Even today, systems that rely on Artificial Intelligence and Machine Learning support people in making important decisions, they can for example help in the selection of personnel. In this case, human-oriented ML means that the employees of the human resources department can understand the key decision criteria of the system. The highest possible degree of traceability and transparency in the decision-making process is necessary to ensure that the functioning of Machine Learning systems is compliant with our legal and social norms and values and that decisions are taken without discrimination.
Artificial neural networks with many layers of information processing are often used to answer complex questions. It is hardly possible for human developers and users to penetrate these deep neural networks and understand how results are achieved. The researchers of the ML2R Competence Center want to penetrate and contain the jungle of deep neuronal networks by using additional knowledge. This knowledge can, for example, take the form of physical equations or be represented as a sequence of logical relationships.
Today, however, it is often still the case that human experts have to label the training data, i.e. provide the correct answers. In the future, the selection and preparation of training data should also be as time-saving and largely automated as possible. Another approach is to consider representative individual cases in decision-making processes, such as the selection of personnel, and to make the decision-making process transparent for these. In this way, the scientists aim to identify the most important influencing factors for a specific individual decision.