Fraunhofer Institute for Experimental Software Engineering IESE
Fraunhofer Institute for Experimental Software Engineering IESE

Artificial Intelligence (AI) in IT Solutions and Products

The use of Artificial Intelligence (AI) will change the world of work. But how are such solutions planned, implemented, and integrated? And what exactly does computer science consider Artificial Intelligence to be?

According to a study by the IWF, every second job in Europe will be changed by AI solutions. Powerful hardware has made the application of AI technologies feasible in practice. In addition, there is a growing number of models, tools, and software components that greatly facilitate the application. However, various quality aspects and framework conditions must be taken into account when designing such systems. The risks associated with AI, and especially with very large artificial neural networks, lie in the areas of safety, security, compliance, and ethics. These risks must be minimized with appropriate methods and measures. At the same time, framework conditions apply, which have been given a legal basis with the EU AI Act and the EU Data Act, in particular.

blau, schwarz, grün, abstrakt, ki, machine learning, deep learning
© unsplash.com/George C; Fraunhofer IESE

What is Artificial Intelligence?



The term Artificial Intelligence was coined in 1956 (Dartmouth Summer Research Project on Artificial Intelligence) to describe the research into understanding intelligence and creating intelligent machines. Today, the term “AI” refers to a software system that applies the methods developed in this field of research (e.g., Machine Learning).

The intelligent behavior on which AI is based can be implemented in various ways, e.g., through rule-based reasoning, case-based reasoning, or Machine Learning. The best-known AI systems today (e.g., ChatGPT, Midjourney, AlphaGo) mostly use Deep Learning techniques.

Machine Learning vs. Deep Learning

What is the difference between Machine Learning and Deep Learning?

In the field of Machine Learning, methods are being developed with which a machine learns to solve certain problems on the basis of examples (data). Today, a distinction is made between symbolic methods (based on rules and logic) and sub-symbolic methods (based on artificial neural networks), between different types of learning (“Unsupervised Learning”, “Supervised Learning”, “Self-Supervised Learning”, “Active Learning”, “Reinforcement Learning”), and between classic learning and “Deep Learning”.

Machine Learning methods generate a decision model that makes a decision given a certain input (e.g., classification of images, text generation, etc.). The main difference between classic Machine Learning and Deep Learning lies in the ways in which the data is processed. In classic Machine Learning, problem-relevant features (e.g., object recognition in images) must be determined from the input data in order to train a model. In Deep Learning, this feature engineering takes place automatically (this is also called representation learning). Deep Learning models are usually larger (more complex) than classic Machine Learning models and require more data for training.

Comparison: Machine Learning and Deep Learning

  Machine Learning Deep Learning
Data volume Small Large
Feature engineering jaYes No
Classifiers Many available Few available
Training time Short Long
Hardware Simple requirement GPU & special-hardware
Traceability Easy to almost impossible Very difficult to impossible

Artificial Intelligence in companies

How do you identify and implement the potential of AI in a company? Which AI use cases bring the most benefits?

blau, schwarz, grün, abstrakt, ki, machine learning, deep learning
© unsplash.com/getty-images; Fraunhofer IESE

Some use cases are easier to implement than others, and the individual benefits also vary from company to company. In addition, an appropriate IT infrastructure is required for development and operation. However, with Fraunhofer IESE’s systematic approach, potential candidates can be quickly identified, evaluated, and implemented as prototypes. Examples of use cases in this context are predictive diagnostics, automatic extraction of data from documents, decision support, image analysis, or automation of dialog systems. For example, AI tools for language processing and translation enable seamless communication across language barriers. In the world of data analytics, AI recognizes patterns and trends and transforms raw, unstructured data into meaningful insights. In CRM, AI personalizes communication and helps to understand and respond to customers’ needs and wishes.

View our service portfolio

Assurance and risks of Artificial Intelligence

How to identify and manage the risks of AI?

Systems that use Machine Learning methods are usually not comprehensible to humans in the way they work and exhibit unpredictable, non-deterministic behavior with some inputs. This means that a specific or desired result cannot be guaranteed. By estimating the probability of occurrence of such errors in relation to the input data, the risk of undesired behavior can be quantified. In combination with other measures surrounding the AI component, systems can be made safer and more reliable.

Understanding and mastering the dependability of AI is a central research topic at Fraunhofer IESE. In recent years, various methods for their evaluation and systematic improvement have been developed, e.g., by evaluating uncertainty. The special characteristics in the development of AI-based functions also lead to novel possibilities for attacking their function (cyber security, secure AI). These can already start with compromised training data or prepared input data intended to provoke functional insufficiencies. Our training seminar addresses the aspects to be considered and how to protect yourself against them. Compliance risks and ethical risks exist if results are issued without the relevant framework conditions being checked. These can be a company’s compliance rules, for example, or legal requirements, e.g., in relation to discrimination against individual groups of people. However, such checks can, in turn, become very complex.

AI in software development

How can Artificial Intelligence improve the software development process and support developers?

blau, schwarz, grün, abstrakt, ki, machine learning, deep learning
© unsplash.com/George C; Fraunhofer IESE

Many software development activities can be made more efficient and support developers with the help of AI-based functions:

  • Identification of contradictions and gaps in natural language requirements
  • Comparison of specifications and extraction of individual aspects (e.g., safety or security aspects)
  • Co-engineering and automated creation of designs and implementations (program code)
  • Generation of test data for system validation
  • Project-specific provision of process knowledge / process tailoring
  • Code reviews, testing of software and finding defects

 

Corresponding functions are increasingly being integrated into software tools. However, company-specific solutions can also be added, e.g., with regard to product-specific aspects.

 

Application areas of Artificial Intelligence

Which AI use cases can Fraunhofer IESE support?

Using Artificial Intelligence is generally possible in all industries and in many use cases. Just get in touch with us! 

Expand all Close all