Neural networks are at the heart of many AI applications that are now arriving in everyday business: text recognition, voice assistants, translation and chatbots. Their name refers to the nervous system, which served as a model for their development. Technically, however, they have little in common with a brain. They are mathematical functions with a very large number of adjustable parameters.
How a neural network is structured
A neural network consists of layers of artificial neurons:
- The input layer receives data, such as the pixels of a photo or the components of a text.
- One or more hidden layers process this data step by step.
- The output layer delivers the result, for example a category or a forecast.
Each neuron receives values from the previous layer, weights them, sums them and passes the result on via a so-called activation function. The weights determine how strongly individual connections have an effect.
Learning takes place during training: the network processes an example, its result is compared with the correct solution and the error is calculated. A method called backpropagation determines how much each weight contributed to this error. The weights are then adjusted step by step so that the error becomes smaller. After a very large number of passes, the network represents the patterns in the data.
Networks with many hidden layers form the basis of deep learning.
Typical architectures
- Feedforward networks for simple classification and forecasting tasks
- Convolutional neural networks (CNN) for images, for example in optical quality inspection
- Recurrent neural networks (RNN) for sequences such as measurement series or texts
- Transformers, the basis of modern large language models. They have largely replaced recurrent networks in language processing.
What companies should keep in mind
- Explainability: The decision logic is distributed across millions or billions of weights and can hardly be read directly. Traceable decisions require complementary methods of explainable AI.
- Quality of training data: A network adopts errors and biases from its data.
- Not every task needs a network: For simple, well-structured data, classic methods are often sufficient, cheaper and easier to verify.
- Computing requirements: Training and running large networks require specialized hardware or cloud resources.
How it differs
The comparison with the human brain is a simplification. An artificial neuron is a simple calculation formula, and a neural network has neither consciousness nor understanding in the human sense. It recognizes statistical relationships in the data it was trained on.
View AI training