Lecture 5: Types of AI

Part 1 — Accessible introduction: What types of AI exist? (20 minutes)

When we say “types of AI,” two fundamental perspectives are meant: a functional perspective (what a system can do) and a technological perspective (how it does it). Functionally, one often distinguishes between "weak" or "narrow" AI, which is specialized for narrowly defined tasks, and the hypothetical "strong" or "general" AI, which would achieve general human-like intelligence. Technologically, one distinguishes, for example, between rule-based systems (symbolic methods) and learning systems (statistics and neural networks). (See Britannica; Russell & Norvig.)

A simple analogy helps: imagine you want to build a house. A rule-based system is like a very detailed construction manual in which a fixed rule applies to every case: "If wall humidity > x, then measure A." A learning system is more like an experienced construction manager who recognizes patterns from many completed projects: after many examples he learns which combinations often lead to cracks, even if no single rule describes that exactly. Expert systems are typically rule-based; modern image or speech processing usually uses learning-based methods like neural networks. (See Britannica: Expert system; Goodfellow et al.)

In everyday life we now encounter almost exclusively weak AI in the form of specialized learning systems or combined systems. Examples are autocorrect while typing, recommendation systems in online stores, image recognition in photo apps, navigation routing, or voice assistants. These systems solve a single limited task very well without "understanding" in the general human sense. (See IBM; Britannica.)

Part 2 — Deepening and technical terms (20 minutes)

Let us begin with the distinction between rule-based and learning-based. Rule-based systems work with explicitly coded rules and a knowledge base. A typical structure is: a knowledge base (facts, rules) and an inference engine that draws conclusions. Such systems are explainable because the rules can be traced, but they are sensitive to incomplete knowledge and rarely scale well when the number of possible situations becomes large. (See Britannica: Expert system.)

Learning systems are based on the concept of a model that is learned from data. Central terms are: dataset (training data), model (mathematical representation), training process (adjustment of the model to the data) and generalization (how well the model performs on unknown data). One broadly distinguishes: supervised learning, in which input examples with target values are available; unsupervised learning, which searches for patterns without given target values; and reinforcement learning, where an agent learns by trial and error and feedback (reward). (See Britannica: Machine learning.)

Within the subgroup of learning methods, neural networks—especially deep neural networks (deep learning)—play a prominent role. They consist of many layers of interconnected nodes and are particularly good at capturing complex patterns in images, audio or text. Deep learning has achieved many practical successes in recent years, however often at the cost of reduced transparency: the internal representations are difficult to interpret, and models can fail unexpectedly when confronted with data that deviates significantly from the training data. (See Goodfellow et al.; Britannica: Machine learning.)

In addition to this technological distinction, there is a horizontal classification by capabilities often discussed in textbooks: reactive systems (which only use currently available information), systems with limited memory (which consider past observations), "Theory of Mind"-like concepts (which model other thinking agents) and hypothetical self-aware systems. These levels are more conceptual and serve for classification; the upper level (self-aware) has not been achieved to date and remains a subject of scientific debate. (See Russell & Norvig; IBM.)

On some important risks and limitations in technical terms: Overfitting refers to the phenomenon that a model fits the training data too closely and therefore generalizes poorly to new data. Bias in a model can arise when training data contain systematic distortions; this leads to discriminatory outcomes. Explainability is a research field that develops methods to make decisions of learning systems understandable. With rule-based approaches explainability is often easier to achieve, while it is more difficult for deep neural networks. (See Goodfellow et al.; Britannica.)

Part 3 — Applications, limitations and thought exercises (10 minutes)

Concrete applications you encounter in everyday life: recommendation algorithms (e-commerce, streaming), speech recognition and voice assistants, automatic translation, image and face recognition in photo apps, fraud detection in bank transactions, navigation and traffic forecasting, autocorrect and text prediction. These systems are mostly combinations of rule-based preprocessing steps and learning models for the actual pattern recognition. (See IBM; Britannica.)

Key practical limitations are: 1) data limits — high quality requirements are placed on training data; 2) robustness — models can fail under deviating conditions; 3) explainability and traceability — particularly relevant in safety-critical areas; 4) generalization ability — current systems are typically specialized and do not exhibit broad, human-like intelligence. For many applications a combination of clear rules and learning components is therefore sensible, often supplemented by human review. (See Goodfellow et al.; Britannica; Russell & Norvig.)

To conclude, three short thought exercises for consolidation: First — name a system in your daily life and justify whether its behavior should be considered rather rule-based or learning-based; pay attention to the role of data. Second — consider what consequences might occur if a learning system is trained on data with systematic bias; what quality-assurance mechanisms would you demand? Third — briefly discuss why a system that is very good at many individual tasks is by no means a strong (general) intelligence; what is the difference between high performance in sub-tasks and general problem-solving ability?

Note on open questions and uncertainties: The terms “weak” and “strong” AI are not formally defined uniformly in science; they serve as a useful distinction, but the exact boundary is conceptual. Likewise, the question of which architectural changes would be necessary to move from specialized to general systems remains on the research agenda; there is no consensus on this. (See Britannica; Russell & Norvig.)