Deep learning – Definition and meaning

What is deep learning?

Deep learning is part of the field of machine learning and relies on artificial neural networks with multiple layers to process large data sets and recognise patterns within them. They are characterised by models with numerous hidden layers, which is why they are referred to as "deep" learning. These network-based architectures can handle a wide range of tasks: From recognising speech and analysing images to generating text or music, the process is used in a wide variety of application contexts. The basic idea is that machines, like humans, should be able to build up and expand their knowledge independently through experience and successive training.

Basic principle and mode of operation

At its core, deep learning works with complex neural networks, in particular deep neural networks (DNNs). Such networks consist of several processing layers, with each layer abstracting specific features. In image analysis, for example, this results in a hierarchy: while the lowest layer recognises edges or simple patterns, subsequent layers recognise more complex structures such as shapes or specific objects, right through to complete objects or facial features.

The learning process is based on training with large amounts of data. Repeated traversal of the network leads to the connection weights of the artificial neurons being gradually adjusted - the aim is to continuously reduce errors in classification or prediction. Methods such as backpropagation and the targeted use of GPUs enable modern deep learning architectures to efficiently process even very large data sets and continuously optimise the models.

Typical areas of application

• Image recognition: Deep learning algorithms identify objects, faces or medical anomalies in photos and videos, for example. One example is the analysis of skin changes to support the diagnosis of skin cancer.
• Speech recognition & processing: Voice assistants such as Siri or Alexa use deep learning models to precisely understand spoken language and trigger appropriate reactions or actions.
• Autonomous driving: Vehicles process sensor data - from cameras, for example - with the help of deep learning, recognise traffic elements and make control decisions for automated driving tasks on this basis.
• Recommendation systems: Streaming services or online shops analyse the behaviour of their users in order to generate individual recommendations for films or products, for example, using deep learning methods.
• Text generation: Artificial intelligence can not only compose or summarise messages, but also create dialogues for chatbots or personalised content.

Examples and scenarios from practice

In the medical field, deep learning-based assistance systems support radiologists in analysing X-ray images and highlighting potential abnormalities. This automated support helps to reduce diagnostic errors and can also shorten the reporting time. In digital commerce, providers rely on deep learning models to check payment and usage behaviour for suspicious patterns almost in real time and identify attempted fraud at an early stage.

Companies without their own data or AI department can also utilise deep learning. Pre-trained models, for example for image or text classifications, can be used to implement targeted applications - such as personalised product recommendations in online shops, which can have a direct impact on the conversion rate. Such solutions often do not require in-depth knowledge of AI development and therefore significantly lower the barrier to entry.

Advantages and challenges

Deep learning approaches offer these advantages in particular:

• Precise results for challenging tasks such as recognising images or patterns in speech signals.
• Automatic extraction of relevant features, which reduces the pre-processing effort for raw data.
• Ability to uncover relationships and patterns hidden in large and complex data sets.

The advantages are offset by various challenges:

• Huge demand for training data and powerful hardware, especially GPUs for model training.
• Low transparency of decisions; deep learning models are often considered "black boxes" that are difficult to understand.
• The effort involved in preparing, selecting and correctly labelling suitable training data should not usually be underestimated.

Recommendations for getting started

Anyone looking to get started with deep learning will benefit from accessible frameworks such as TensorFlow or PyTorch, for which numerous tutorials and pre-trained models exist. These offer practical access to typical problems such as sorting images or classifying text. After initial success with standard models, it is advisable to take a look at the data basis: The quality and sensible preparation of the training data has a significant influence on the performance and informative value of even sophisticated deep learning projects.