Page 28 - CyberSecurityEssentialsEnglish

Hybrid AI systems combine learning from data with rules established by experts. This
approach is useful when we want the flexibility of a model while enforcing boundaries that

must not be crossed, such as safety rules, legal requirements, or organizational policies.
Early AI applications relied on rules, logic, and expert systems. As data and computing power

grew, the field expanded to include machine learning and deep learning, supporting major
advances in image, audio, and language applications.

No single approach suits every problem. If a task is governed by a clear, stable rule, a rule-
based program may be simpler and easier to understand. If the task requires the detection
of numerous complex patterns, learning from data may be more appropriate.

The newest approach is therefore not always the best. The best approach is the one that fits
the problem, available data, and required levels of accuracy and transparency.
Within machine learning, deep learning uses neural networks built from many successive

layers. Each layer transforms the data and passes the result to the next, so the model can
move from simple features to more complex ones without a person specifying them in
advance. The word deep refers to the number of these layers, not to depth of understanding.

Growth in data and computing power made this approach practical, and it lies behind most
modern applications in image, audio, and language.

Generative AI refers to systems that produce new content, such as text, images, audio, video,
or computer code, rather than only classifying, ranking, or predicting. Most current

generative systems are built on deep learning. Producing new content, however, is not the
same as retrieving a verified fact, which is why the output of a generative system must be
checked against a reliable source.

A model of this kind, trained broadly and then adapted to many different tasks, is often called
a foundation model. The name describes its position rather than its architecture: it is a

general-purpose starting point that other applications are built on, instead of a system
trained from scratch for one job. This is why the same underlying model can sit behind a
writing assistant, a classification tool, and a customer service system at the same time, and

why a change to that model can affect all of them at once.

27