General Definition

It is difficult, if not impossible, to define the concept of artificial intelligence in a single sentence. Even within the scientific community, no single definition has gained widespread acceptance. This is not surprising, since human intelligence itself cannot be clearly defined. Here are a few common definitions:

Artificial intelligence is the ability of a computer or computer-controlled robot to perform tasks that are commonly associated with the intellectual processes characteristic of humans, such as the ability to reason. Although there are as of yet no AIs that match full human flexibility over wider domains or in tasks requiring much everyday knowledge, some AIs perform specific tasks as well as humans.

source: https://www.britannica.com/

The capability of a machine to imitate intelligent human behavior, or the theory and development of computer systems able to perform tasks that typically require human intelligence.

source: https://www.englishclub.com/

Artificial intelligence (AI), is a term coined in 1955 by John McCarthy, Stanford’s first faculty member in AI, who defined it as “the science and engineering of making intelligent machines.” Much research has human program software agents with the knowledge to behave in a particular way, like playing chess, but today, we emphasize agents that can learn, just as human beings navigating our changing world.

source: https://hai.stanford.edu/

Strong AI and Weak AI

Imagine there’s a robot that schedules appointments for you, handles annoying calls, cleans up the mess in your apartment—all while making sure you still know your way around, don’t lose track of things, and know where everything is. Plus, this robot is always your best friend; you can tell it anything, and it understands me, listens to me, and gives me what I need.

Unfortunately, that’s still a long way off. The AI we have available today can, unfortunately, only handle the specific tasks it has been trained for. This brings us to the difference between strong and weak AI.

Weak AI, also called narrow AI, is capable of performing a specific task that it’s designed to do. Strong AI, on the other hand, is capable of learning, thinking and adapting like humans do. That said, strong AI systems don’t actually exist yet.

source: https://www.britannica.com/

Strong AI and Weak AI

Here are the typical characteristics of strong and weak AI, as well as a comparison of cases where the term “AI” does not apply at all.

Strong AI and Weak AI

Of course, there is a wealth of additional useful information on the topic of AI available online. We have compiled and linked some helpful resources for you here.

Useful Video:

Useful Link:

Subfields of AI

Many of us think of AI as traditional chatbots like Chat GPT. In reality, the field of AI encompasses many more subfields. We likely only come into indirect contact with many of these areas in our daily lives and are therefore unaware of them. Here is a simple diagram showing key aspects of AI. It is highly simplified, by no means exhaustive, and is intended only to provide an overview.

Computational Intellicence (CI)

This is the main subset of artificial intelligence (AI) that draws inspiration from natural intelligence. It is capable of acquiring knowledge and adapting to complex situations by mimicking human-like reasoning and learning processes.

Neuronal Networks

The field of artificial neural networks has emerged as a significant area of research in recent years. Artificial neural networks mimic the structure and function of biological neural networks. The next section will cover neuronal networks.

Other...

There are many other areas that fall under the umbrella of AI and CI. The average user usually encounters these only indirectly. The diagram on the right shows just a selection of them. It is intended to illustrate the scope and diversity of the term AI. However, the following discussion will not delve further into the subfields outside of neural networks.

More Information

The links below offer more in-depth and useful information on the subject of artificial intelligence.

Useful Links:

General Definition

Strictly speaking, when discussing neural networks in the context of artificial intelligence, it is more accurate to refer to them as artificial neural networks. This is because there are also neural networks in biology:

Biological neural networks (BNNs) are large, interconnected systems of neurons found in living organisms, including the human brain. They process information through electrochemical signals and form the base for thinking, memory, emotion, and movement.

source: https://www.upgrad.com/blog/biological-neural-network/

Artificial neural networks attempt to mimic these biological networks, such as those in the human brain. They can be defined as follows:

A neural network is a machine learning model that stacks simple "neurons" in layers and learns pattern-recognizing weights and biases from data to map inputs to outputs.

source: https://www.ibm.com/think/topics/neural-networks

Neural networks are machine learning models that mimic the complex functions of the human brain. These models consist of interconnected nodes or neurons that process data, learn patterns and enable tasks such as pattern recognition and decision-making.

source: https://www.geeksforgeeks.org/deep-learning/neural-networks-a-beginners-guide/

Structure of an Artificial Neuronal Network

An artificial neural network consists of multiple layers. There is at least one input layer and one output layer. Additionally, there can be any number of hidden layers in between. Each layer consists of nodes, known as neurons. These neurons are connected to one another via weighted connections. Weights are numbers that can be arbitrarily large, small, or even negative.

The image on the left shows a classic artificial neural network. (The weights are not shown in the graph on the right.) This network has one input layer, one output layer, and two hidden layers. Each neuron is connected to every neuron in the next layer. This type of neural network is also known as a fully connected network. It has become the standard for depicting neural networks. However, there are artificial neural networks with more complex structures. For example, connections may be missing or span multiple layers.

The neural networks used in practice are much larger than the one shown in the image on the right. They can consist of several hundred layers and contain billions of neurons. Conversely, smaller networks are also used that have fewer than ten layers and approximately ten thousand neurons.

Functionality of an Artificial Neural Network

How does an artificial neural network work? As you might expect, the operation of an artificial neural network is not trivial. There is some mathematics behind it. However, it’s not as complicated as it seems at first. A rough explanation is provided below to give you a basic understanding. The next slides include links to help you learn more about artificial neural network.

First, input data is fed into the input layer. This requires converting the input layer into a format that the neural network can read. Text is converted into numbers because a neural network cannot read text. It can only process numbers. Each word is assigned a list of numbers representing its meaning. Images are broken down into pixels, which are then converted into numbers. Each neuron in the input layer receives the numerical value corresponding to a pixel in a specific position.

The input data is processed through subsequent layers. Each neuron receives inputs from other neurons. Then, a calculation is performed based on

• the inputs, which are received by the neuron
• the weights, which are assigned to the connections
• the bias, which is assigned to the neuron
• and the activation function, which is assigned to the neuron.

These components determine the activation value, which is simply a number. The calculation is not complicated. First, all weighted inputs are summed up. Then, the bias is added. Finally, the activation function is applied to the result. The activation function can vary greatly. It is what enables the artificial neural network to learn complex patterns.

The figure shows the calculation for the activation value of a single neuron. x represents the inputs that the neuron receives from other neurons. w represents the weights. These indicate how much influence each input has on the activation value. The neuron has an activation function, A, and a bias, b. The activation value is given by

and serves as the input for the next neurons. This process continues through all the layers in this way. The activation functions, executed in the final layer of the neural network, can be considered output functions. These functions produce numerical values that must be converted into a readable format.

Functionality of an Artificial Neural Network

How does an artificial neural network learn? The previous slide described how an artificial neural network works and mentioned terms like "weights" and "bias." But where do these values come from? For large artificial neural networks with several hundred layers, the number of weights is too great for humans to assign manually.

Initially, the artificial neural network is randomly initialized. It therefore consists of randomly generated initial weights that generally do not make sense. After that, training begins. During this time, the weights and biases are adjusted as much as possible for the intended tasks.

Training data is used for this purpose. This data includes tasks similar to those that will be encountered later on in practice. For example, if you have an neural network designed to recognize faces in photos, you would train it using real photos. With training data, it is important that the results can be evaluated. This means you must be able to provide the neural network with feedback indicating whether it has performed a task correctly or incorrectly.

First, present the artificial neural network with its first task. The network responds using the randomly initialized weights. The first answer will therefore be incorrect. Then, the neuronal network receives feedback indicating that it was wrong. Subsequently, the neuronal network adjusts its weights and biases to minimize the error. So far, the neuronal network knows only one input value. Training with the next input will therefore fail again. However, an adjustment is made there as well. This process is repeated millions of times. With each error, the network receives clear guidance on how to improve. As a result, its accuracy increases step by step.

What about the activation functions? Are they trained as well? No, they are predefined, as are the number of layers and neurons.

Is the output of an Artifical Neuronal Network Random?

Take a moment to review the previous two slides on how an artificial neural network works. You’ll see that it’s a simple function that follows simple rules. The complexity of a neural network stems from its enormous size. Since a neural network (NN) is defined by precise rules, shouldn't it be deterministic? But what exactly does "deterministic" mean?

Determinism, in philosophy and science, the thesis that all events in the universe, including human decisions and actions, are causally inevitable. Determinism entails that, in a situation in which a person makes a certain decision or performs a certain action, it is impossible that he or she could have made any other decision or performed any other action. In other words, it is never true that people could have decided or acted otherwise than they actually did.

source: https://www.britannica.com/topic/determinism

Deterministic refers to a process or function that operates in a predictable manner. In both programming and mathematics, deterministic systems yield the same output from a given set of initial conditions without any randomness involved.

source: https://codingtechroom.com/question/what-does-deterministic-mean

An artificial neural network would therefore have to be deterministic. So if I made a copy of a chatbot and asked both exactly the same questions, I would have to receive exactly the same answers.

In theory, yes. However, chatbots used in real-world applications often incorporate additional random components.

More Information

The links below offer more in-depth and useful information on the subject of neuronal networks.

Useful Links:

Useful Videos:

Large Language Models (LLM), Machine Learning (ML) and Deep Learning (DL)

On the last few slides, we saw that the artificial intelligence (AI) we use every day is made up of artificial neural networks (ANNs). In the context of AI, the terms "large language models", "deep learning" and "machine learning" come up time and again. Are these simply other terms for neural networks? No, not exactly. These terms actually all have different meanings. The following section clarifies how they can be categorized in relation to AI and ANNs.

Machine Learning (ML): The first term that we will examine in more detail is 'ML'. It can be defined as follows:

Machine learning is the subset of artificial intelligence (AI) focused on algorithms that can “learn” the patterns of training data and, subsequently, make accurate inferences about new data. This pattern recognition ability enables machine learning models to make decisions or predictions without explicit, hard-coded instructions.

source: https://www.ibm.com/think/topics/machine-learning

This technical definition is not easy to understand. Put simply, however, ML is the subfield of AI in which algorithms learn from data. Thus, ML enables algorithms to make predictions and decisions without explicit programming. ANNs are a subfield of ML, but ML encompasses much more than just neural networks. Machine learning also includes statistical methods such as regression.

Deep Learning (DL): Another term that frequently comes up in the context of AI is DL, which can be defined as follows:

Deep learning is a subset of machine learning driven by multilayered neural networks whose design is inspired by the structure of the human brain. Deep learning models power most state-of-the-art artificial intelligence (AI) today, from computer vision and generative AI to self-driving cars and robotics.

source: https://www.ibm.com/think/topics/machine-learning

DL is therefore a subcategory of ML that uses neural networks. It's important to note that these networks have multiple hidden layers. These layers enable the networks to recognize complex patterns in large, unstructured datasets. The term "deep" in DL refers to the deep hidden layers of the neural network.

Large Language Models (MML): And, of course, we mustn’t forget one final term: MML, which can be defined as follows:

Large language models (LLMs) are a category of deep learning models trained on immense amounts of data, making them capable of understanding and generating natural language and other types of content to perform a wide range of tasks. LLMs are built on a type of neural network architecture called a transformer which excels at handling sequences of words and capturing patterns in text.

source: https://www.ibm.com/think/topics/machine-learning

From this perspective, LLMs are deep learning models on a much larger scale. The transformers mentioned in the definition enable parallel processing. A more detailed explanation of transformer architecture would exceed the scope of this training.

The above picture shows how all terms are classified and categorized: AI (Artifical Intelligence) - CI (Computational Intelligence) - ML (Machine Learning) - ANN (Artifical Neuronal Networks) - DL (Deep Learning) - LLM (Large Language Models).

Large Language Models (LLM), Machine Learning (ML) and Deep Learning (DL)

Here is a summary, along with some additional helpful details about the terms ML, DL and LLM.

More Information

The links below offer more useful information abeut the various terms we repeatedly encounter when dealing with AI.

Useful Links:

Useful Videos:

Strength of Neuronal Networks

Neural networks are capable of:

• processing large amounts of data,
• identifying patterns within that data,
• can learn from that data without requiring external adjustments.

This makes them highly versatile, and we have likely not yet fully tapped into their potential or explored all their applications. The use of neural networks is still in its infancy, especially in fields such as medicine, research, and finance. Humans cannot process such large amounts of data, and traditional algorithms quickly reach their limits in this context. Neural networks can recognize patterns where humans see only "coincidence."

Nevertheless, even though neural networks are capable of tremendous feats, we shouldn’t rely on them blindly. You’ve probably already encountered an incorrect or illogical response when working with LLMs. This brings us to the next topic: the weaknesses of neural networks.

Weakness of Neuroanlal Networks ans especially Large Language Models

Of course, distortions and incompleteness can occur during the design process and, more importantly, during training with artifical neuronal networks (ANN), leading to undesirable behavior. Below are a few of these issues and other drawbacks:

• Resources: Large networks require powerful hardware to train efficiently. This consumes a great deal of resources and has associated environmental impacts.
• Lack of data: ANN require large amounts of data to truly learn. Such large datasets are not available in all fields.
• Overfitting: This occurs when the network adapts too closely to the training data. Consequently, it becomes difficult to generalize the model to new data.
• Interpretability: Deep models are difficult to interpret, so it is impossible to understand exactly how the neural network arrived at its output.
• Transferability: The same model may not perform as well in a different domain.
• Causality: ANNs can detect correlations well but not necessarily causality.
• Robustness: ANNs may respond poorly to unexpected inputs that they were not trained on.
• Data bias: Biases in the data can lead to unfair predictions.
• Data privacy: Models can extract sensitive information from the data

Most of the points listed refer not to the neural network itself, but to the training data. An ANN is only as good as the data used to train it. For example, if a neural network is only fed images of cows and pigs, it will not be able to recognize a lion. This is especially critical if an ANN is trained only with data from men and then fails to account for aspects that particularly affect women. This is especially true in sensitive fields, such as medicine.

To learn more about this important topic, the next slide contains helpful—and, in some cases, entertaining—links.

More Information

The links below offer more useful information abeut the various terms we repeatedly encounter when dealing with AI.

Useful Links:

Useful Videos:

It helps you better understand what AI can and cannot do. For example, you’ll realize that AI doesn’t have a “mind of its own,” but rather, it merely recognizes patterns in data. You’ll find it easier to understand why it’s sometimes right and sometimes wrong. This will prevent you from blindly believing everything. You'll be better able to evaluate results, protect yourself from manipulation, and use AI more effectively.

Plus, you’ll stay ahead of the curve. AI is becoming the standard in many fields. Those who understand how it works will have an easier time at work and in everyday life.

Of course, if there’s technology that can perform tasks we dislike, we should use it, even if it isn't perfect. After all, that applies to almost all technologies. The important thing is not to rely on it blindly, but rather to learn when to question the results. You probably already do that when you ask a friend for help, though.

Chatbots, for example, are like friends who know a lot but sometimes struggle with logical thinking. They always want to please you and agree with you more often than is realistic. Caution is advised here, too.

However, if you never use new technologies, you'll never learn how to work with them.

Unfortunately, there’s no clear answer to this question. For example, do you find automatically filled-in calendars annoying? This results in you receiving appointments that aren’t relevant to you. Since you didn't enter the appointments yourself, you lose track of them. This means you have to check your calendar much more often, creating more work for you overall.

AI can save you work, but it doesn't always.

In your personal life, at least, you have the choice of how, when, and to what extent you want to use AI. Figure out what works best for your needs.

You are right, artificial intelligence can be described as a system that recognize patterns, makes decisions, or learns from data.

Unfortunately, you answer was not correct. The right answer is: artificial intelligence as a system that recognize patterns, makes decisions, or learns from data.

You are right, Google maps, recommendations from Amazon and a spam filter can be classified as weak AI. The superintelligence that will turn humans into slaves and Data from Star Trek are strong AI. A calculator or The Super Mario World game cannot be classified as AI.

Unfortunately, you answer was not correct. The right answer is: Google maps, recommendations from Amazon and a spam filter can be classified as weak AI. The superintelligence that will turn humans into slaves and Data from Star Trek are strong AI. A calculator or The Super Mario World game cannot be classified as AI.

You are right. Neuronal Networks, Computational Intelligence and Swarm Intelligence are Subfields of AI.

Unfortunately, you answer was not correct. The right answer is: Neuronal Networks, Computational Intelligence and Swarm Intelligence.

You are right, chatbots can be classified as Neuronal Networks.

Unfortunately, you answer was not correct. The right answer is: chatbots can be classified as Neuronal Networks.

You are right: an artificial neuronal network must have at least one input layer; the brain of a mice is a biological neuronal network; the weights in a neural network can also be negative

Unfortunately, you answer was not correct. The right answer is: an artificial neuronal network must have at least one input layer; the brain of a mice is a biological neuronal network; the weights in a neural network can also be negative

You answer was correct. Diagram D does not correctly describes the structure of a neural network?

Unfortunately, you answer was not correct. The right answer is: Diagram D does not correctly describes the structure of a neural network?

You answer was correct. Machine learning is not a subset of Deep Learning; The construction of Large Langue Models (LLM) always involves the use of a neural network; "Deep in deep learning" is a term used to describe the multiple "deep" layers of a neuronal network.

Unfortunately, you answer was not correct. The right answer is: Machine learning is not a subset of Deep Learning; The construction of Large Langue Models (LLM) always involves the use of a neural network; "Deep in deep learning" is a term used to describe the multiple "deep" layers of a neuronal network.

You answer was correct. The two answers will most likely be different.

Unfortunately, you answer was not correct. The right answer is: The two answers will most likely be different.

You answer was correct. This statement is false.

Unfortunately, you answer was not correct. The right answer is: This statement is false.

The responses will be sent to the email address of your account. If you wish to receive your answers to a different address, you can change it in the field below.