Deep Learning.

Deep learning is a method of machine learning in which artificial neural networks with many layers learn from examples. According to IBM, deep learning differs from simpler neural networks in that it has multiple hidden layers; IBM cites four layers as the minimum. Almost everything that is considered AI today is based on it: large language models, image recognition, speech recognition, and image generators. For your company, the term is particularly useful for understanding what an AI solution can do, what it requires, and where its limitations lie.

How Deep Learning Works

A neural network learns by adjusting its internal weights until the outputs are correct. According to IBM, the upper layers detect coarse patterns, while the deeper layers detect increasingly finer ones. The network thus learns on its own which features in the data are important, rather than having a human specify them.

According to IBM, deep learning typically requires very large amounts of training data and significant computing power for both training and deployment. It was only with the advent of graphics processing units—which perform many calculations in parallel—that this became feasible on a large scale. According to IBM, large language models typically use what is known as the Transformer architecture, while image recognition often relies on what are known as convolutional networks.

What You Should Keep in Mind at Work

Don't train your own model if a ready-made one will suffice. Because data and computing power are so costly, most companies use existing models via an interface. For your own content, it's often enough to provide the model with appropriate documents—see RAG—or to fine-tune it for better performance.

Expect results that cannot be fully explained. IBM describes deep learning models as a “black box”: The learned parameters provide little insight into how they relate to the data. Therefore, whenever you need to justify a decision—for example, to customers or job applicants—you’ll need a transparent rule or a human review.

Check your data before you start thinking about building your own model. In AI projects, we often see that the idea of building a custom model falls through because of data limitations: too few examples, inconsistent labeling, or data scattered across many systems. According to IBM, labeling large datasets is particularly expensive.

Frequently Asked Questions About Deep Learning

What is the difference between machine learning and deep learning? Deep learning is a subset of machine learning. According to IBM, traditional machine learning often relies on features selected by humans, whereas deep learning learns the features from the raw data itself.

Is generative AI also deep learning? Yes. According to IBM, generative AI and large language models are among the most important applications of deep learning. Learn more under Generative AI.

Does our team need to understand deep learning to use AI? Not in detail. What’s important is a basic understanding of its strengths and limitations—for example, why models can sound convincing yet still be wrong. That is precisely what AI competence under Article 4 of the AI Act is all about.

If you want your team to understand what AI tools are all about and where their limitations lie, check out our AI workshop.