The AI Act Handbook

Gabriele Bolek-Fügl

Artificial intelligence is at the center of a revolution that is not only changing the way machines think and learn, but also challenging our understanding of intelligence itself. This revolution is driven by big data and the extraordinary computing power of modern computers, which can perform complex statistical calculations with a speed and precision that has long been unimaginable.

AI has many faces. From simple programs that perform individual tasks with an accuracy that surpasses human ability, to complex systems that learn to adapt and make decisions that are reminiscent of the human mind.

Increasing digitalization can free us from repetitive and tedious tasks and give us the freedom to act more creatively and strategically. If you have little talent or experience in one area, you can compensate for this with virtual assistants. AI offers opportunities to expand our skills and push our boundaries. However, to take full advantage, we need to learn how to work effectively with AI.

But what exactly makes a machine "intelligent"? Is it its ability to recognize and respond to human speech? Its efficiency to make decisions in milliseconds that a human could only make with difficulty and after long deliberation? Or is it the ability to learn from experience and improve over time?

In fact, AI encompasses a spectrum of technologies that are as diverse as the definitions that seek to capture them. So let's begin our journey into the world of AI with a look at the basic components and the various techniques that are summarized under the term "artificial intelligence".

In the field of AI, names and terms are often used without knowing the exact definitions and backgrounds. But in the world of mathematicians, computer scientists and architects of AI, precision is essential. Clear, unambiguous definitions are needed to push the boundaries of what is possible and develop the next generation of intelligent systems.

When developing any AI system, there are components that are necessary regardless of the specific technology or use case. These cornerstones form the foundation on which more complex AI algorithms can be built. Let's therefore take a look at these components of AI-systems and define the associated details:

Tabelle 1.1 Overview of important AI Components

Component

Description

Importance

Complexity

Customizability

Data

Information for learning and decision-making

High

Variable

High

Algorithms

Procedures or methods for processing data

High

High

High

Computing power

For processing large amounts of data and complex calculations

High

Medium

Medium

Storage

Necessary for storing data, models and results

Medium

Low

Medium

Measurement and model optimization

Important for evaluating the effectiveness of AI models and their optimization

High

High

High

Interfaces for interaction

Enable interaction between humans and AI systems

Medium

Medium

High

Security and data protection

Protect data from unauthorized access and misuse, ensure compliance with legal standards

High

High

Medium

Importance: A rating of "High" means that this component is very important in the context of the correct and efficient operation of an AI system. Medium and Low indicate less important criteria.

Complexity: The rating refers to the degree of difficulty with which reliably correct or performant results are generated by the AI.

Adaptability: The rating indicates how flexibly a company can adapt the components of an AI to its own needs and tasks.

Let us now take a closer look at the individual components. The following explanations represent only the most important aspects and do not claim to be exhaustive.

Data is the fuel for AI models. Similar to a vehicle, the AI model would be of no use without data. It provides the information needed to recognize patterns, make decisions and learn from experience.

Data is the lifeblood of AI. Without data, AI would have no point of reference to analyze, learn or improve. The quality of the data used is therefore of particular importance! If the quality is not sufficiently high, it will not be possible to achieve good results.

The availability and quality of data has steadily increased in recent years. The development of the internet and digital technologies has led to an exponential increase in the amount and variety of data that AI systems now use for learning and analysis.

Data must always be in context of its use in order to understand its meaning and value. Context provides insight into what data represents, the context in which it was collected and how it should be interpreted. For example, depending on the context, the number 9 could represent a temperature (in degrees Celsius), a length (in centimeters), a duration (in minutes) or a variety of other measurements.

In data analysis and artificial intelligence, the context is crucial for drawing correct conclusions and making accurate predictions. Models require training with data that is not only accurate and cleansed, but also presented in a suitable framework. The context therefore directly influences the outcome of an AI model:

Tabelle 1.2 Goals and Challenges for the Context of AI Result Generation

Range

Goal

Challenge

Example

Data selection

Selection of data that is relevant and representative of the problem.

Insufficient or distorted data can lead to poor model performance.

Selection of patient data for an AI model to predict diabetes.

Feature engineering

Development and transformation of features to improve the presentation of information for the model.

Overfitting to training data and neglecting generalizability.

Creation of time windows for customer buying behavior in an AI model to predict sales.

Model interpretation

Analyze and explain the results produced by an AI model.

The traceability and verification of the decision-making processes of complex models such as deep neural networks are often difficult to carry out.

Interpretation of the credit rating in a financial model.

Depending on the format in which the data is available, the context can be better or worse recognized and processed.

       Structured data are organized in a predefined format, usually in tabular form, which facilitates their analysis and processing. Examples of this are databases and Excel tables.

       Unstructured data on the other hand, have no predefined format or structure. It includes texts, books, images, videos and more. Processing this data requires more complex methods, as structuring is necessary first.

Extensive and well-prepared data sets are required for the training of AI models. These data sets are divided into three main categories:

1.      Training data is the data on which the model is trained. They form the basis for learning patterns and correlations.

2.      Validation data is used to evaluate the performance of the model during training and to adjust the parameters.

3.      Testing data are used to evaluate the final performance of the model. They must be independent of the training and validation data in order to ensure an objective evaluation.

The quality of the data is also crucial for the performance of AI models. High-quality data is accurate, complete, consistent and up-to-date. Poor data quality leads to unreliable models that deliver distorted or incorrect results. It is often not until the results after the training phase that you realize that the data quality is not sufficient for the intended purpose and then have to rework the data.

Another major challenge is the legally compliant processing of sensitive and personal data as well as data from minorities. Solutions such as anonymization of data and methods to reduce...