Source: Freepik
Neural networks are models inspired by the functioning of the human brain. They consist of artificial neurons that process information and communicate with each other. These networks enable machines to learn from data and recognise patterns without any explicit programming for each task. Examples include virtual assistants such as Alexa and Siri, translation tools such as Google Translate and creative AI systems such as ChatGPT. In this article, we will look at how neural networks work, what advantages and disadvantages they have and how RWTH Aachen University uses them.
How Do They Work?
Each connection between the neurons in a neural network has a so-called weight. This weight determines how strongly a signal is transmitted from one neuron to the next. You can think of it like a volume control: A higher weight amplifies the signal; a lower weight attenuates it.
During the training process of the network, these weights are adjusted to improve the performance of the model. The network compares its predictions with the actual results in a so-called training data set.
A concrete example:
- Let’s assume that the network is supposed to differentiate between images of cats and dogs.
- At the beginning, the weights are set randomly, and the network makes inaccurate or incorrect decisions.
- An algorithm is used to calculate how big the prediction error was.
- This error is used to adjust the weights so that the network becomes more accurate in future runs.
The process of weight adjustment is what makes the network ‘learn’. Over time, the network will be able to recognise patterns in the input data (for example the shape of the ears or the texture of the fur) and make reliable decisions based on this.
How Are They Structured?
A neural network consists of several layers:
- Input layer: This is where the data is recorded in numerical form. The data can be images, texts, numbers or sound recordings, for example. It often comes from external sources, such as databases, sensors, cameras or user interactions, and is often processed or adapted before being entered into the network in order to make it usable for the analysis. For example, pixel values of an image or words of a text could be processed.
- Hidden layers: These layers contain the actual computational processes of the network. They analyse the data, recognise patterns and make decisions. Each connection between the neurons has a so-called weight, which is adjusted during training.
- Output layer: The result (for example the prediction or classification) is provided here. This layer is responsible for interpreting and passing on the results to the user or other systems. Depending on the use case, the result can be a number, a category or even a complete text. The accuracy and usefulness of the output layer depends heavily on the quality of the data and the efficiency of the previous processing in the hidden layers.
However, it is important to note that neural networks do not always deliver correct results. Especially if the input data is incorrect, incomplete or unsuitable, incorrect outputs can occur. The training process therefore plays a central role: it helps the network to learn from many examples and deliver more precise results. Nevertheless, the results are highly dependent on the quality of the data and the training methodology.
Where Are Neural Networks Used?
We encounter neural networks in many areas of everyday life:
- Image recognition: For example, in facial recognition on mobile phones (FaceID) or medical image analysis.
- Voice processing: Applications such as translation tools or voice assistants like Alexa or Siri use neural networks.
- Autonomous systems: Self-driving cars and drones use them to understand their surroundings.
- Personalisation: Recommendation algorithms in social media or streaming services are also based on neural networks.
Challenges and Future Prospects
Despite their versatility, there are challenges:
- Data hunger: neural networks require large amounts of training data.
- Computational effort: Training is often time-consuming and resource-intensive.
- Transparency: Their decisions are often difficult to understand, which can be problematic for sensitive applications. This is because they perform complex mathematical operations in several hidden layers that cannot be easily interpreted by humans. This lack of transparency can be problematic for sensitive applications, such as in medicine or finance. For example, you might want to know why an AI system has made a certain diagnosis or rejected a loan application.
To tackle this challenge, researchers are working on concepts such as ‘Explainable AI’ (XAI). This approach aims to make the decision-making processes of AI systems more understandable and comprehensible. Explainable AI uses techniques that show which features of the input data (for example certain pixels in an image or words in a text) were particularly important for the network’s decision. An example of this is the visualisation of activations in the layers of the network that show which areas of an image the model has classified as relevant.
Neural Networks at RWTH
RWTH Aachen University is one of the pioneers in the integration of artificial intelligence into everyday university life. With the RWTHgpt project, which has been available since July 2024, the university is using generative AI technologies such as GPT-4o, GPT-o1 and soon GPT-o3 to set new standards in research, teaching and administration. You can read more about this in our introductory blog post. You can also find another article that specifically addresses availability for students. The AI tool was developed in collaboration between the IT Center and the Center for Teaching and Learning Services (CLS) and enables users to set text-based tasks and analyse multimodal content. Particular emphasis is placed on data protection, meaning that all data is processed exclusively on servers in the EU. The use of RWTHgpt is data protection-compliant and enables fast and secure access to the latest AI models, which offers considerable added value for both employees and students.
The integration of AI at RWTH Aachen University shows how modern technologies can be used to optimise existing processes and create new opportunities. Training materials and how-tos for the targeted use of generative AI systems for teaching, learning and testing are available at the RWTH LehrBar. Through RWTHgpt in research, teaching and administration and through other AI systems, RWTH focuses on innovative approaches to master challenges efficiently. With a clear focus on data protection and user-friendliness, the university remains a pioneer in the use of AI.
Responsible for the content of this article is Malak Mostafa.
The following sources served as the basis for this article:
Leave a Reply