Learn
6.3.2 | CHARACTERISTICS OF AI
01 | THE CHARACTERISTICS IN THE SPECIFICATION
The specification describes AI through the collection of data and rules for using that data, the ability to reason, and the fact that it can include the ability to learn and adapt.
These ideas describe simulated intelligent behaviour, not a claim that the computer has human thoughts or feelings.
Can include is important: learning is not required in every AI system. An expert system can reason using fixed knowledge and rules.
02 | DATA: WHAT INFORMATION IS AVAILABLE?
Data supplies the information an AI system uses. It can come from user answers, records, sensors, images or other appropriate sources.
| System | Possible data |
|---|---|
| Printer troubleshooting expert system | User answers about power, paper and error indicators |
| Image classifier | Images and labels during training; a new image during use |
| Prediction system | Relevant past observations and new input values |
Collecting data is not itself AI. The method used to interpret it and produce a result matters.
Relevant, accurate information is more useful than simply a large amount. Missing, wrong or unrepresentative information can reduce the quality of conclusions.
03 | RULES: HOW WILL THE DATA BE USED?
A rule specifies how information leads to a conclusion or action. In an expert system, rules can be written explicitly using expert knowledge.
Teaching rule: IF the printer has power AND paper is absent THEN recommend checking the paper tray.
The system checks the stated conditions against the available facts. If they match, the corresponding conclusion can follow.
A rule may combine several facts. It is not a random answer and does not require the computer to have personal experience.
04 | REASONING: CONNECT FACTS TO A CONCLUSION
Reasoning means using information and relationships or rules to reach a conclusion.
| Step | Teaching expert-system example |
|---|---|
| Available facts | Power is present; paper is absent |
| Relevant rule | If power is present and paper is absent, recommend checking the tray |
| Check | Both stated conditions match |
| Conclusion | Recommend the paper-tray check |
The result is conditional on the facts and rule being appropriate. It does not prove that the paper tray is the only possible fault.
The AI characteristic is the simulated reasoning process. Merely storing the answers in a file does not by itself explain that process.
05 | FOLLOW THE INFORMATION PATH
COLLECT DATA
Obtain facts or examples relevant to the task
↓ Use a defined method
REASON OR INFER
Apply rules or learned relationships
↓ Produce a conclusion
OUTPUT AND CHECK
Recommendation, classification or prediction
This is a simplified view, not a requirement that all AI algorithms have identical internal steps. The next lesson examines specific expert-system and machine-learning components.
06 | LEARNING: CHANGE USING EXPERIENCE OR DATA
In machine learning, a program can adapt its processes or data using training information. Training changes a model so it can perform a task based on patterns in the data.
Example: labelled material images can be used to train a classifier. The model is adjusted during training to distinguish categories using relevant features.
This is different from a person manually writing a separate rule for every image. People still choose the task, training method, examples and evaluation criteria.
Learning is a computational adaptation process; it does not establish human understanding.
07 | ADAPTATION: WHAT ACTUALLY CHANGES?
Adaptation is a change in the system’s model, processes or data that affects later behaviour.
| Situation | What changes? |
|---|---|
| Training a classifier | Model values are adjusted using examples |
| Updating a rule-based system manually | A person changes its knowledge or rules |
| Using a fixed trained classifier | Input and output change, but the model may remain unchanged |
A different input can produce a different output even when the method is fixed. Do not call every new answer evidence of adaptation.
Automatic learning differs from a programmer editing the system by hand, although both can change behaviour.
08 | TRAINING AND USE ARE DIFFERENT
Training: information is used to adapt the model.
Use, or inference: the trained model processes a new input to produce a result.
A model can be trained first and then deployed without changing during ordinary use. Some systems support further or ongoing learning, but this is not universal.
Performance should be checked using suitable data not used to fit the model. A system that performs well on familiar training examples may still fail on unfamiliar cases.
09 | COMPARE THE TWO COURSE APPROACHES
| Characteristic | Expert-system example | Machine-learning example |
|---|---|---|
| Data | Facts and user answers | Training examples and new inputs |
| Method | Explicit knowledge and rules | Relationships represented by an adapted model |
| Reasoning / inference | Apply rules to reach a conclusion | Use the model to produce a prediction or classification |
| Learning | Not required for a fixed rule-based system | Training adapts the model |
| Change during ordinary use | Not necessarily | Not necessarily |
This is the comparison used in this course. Real systems can combine rule-based and learned components.
10 | QUALITY, BIAS AND LIMITS
Wrong input: an expert system can reach an unsuitable conclusion if a user supplies an incorrect fact.
Incomplete rules: a rule-based system may not cover an unusual situation.
Unrepresentative training: a model may work better for familiar cases than poorly represented cases.
Changing conditions: relationships learned from past data may not remain useful for a new environment.
Learning does not guarantee improvement on every task or every group. Testing and careful evaluation remain necessary.
11 | A WORKED COMPARISON
Two systems help sort recyclable items.
System A: applies an explicit rule to known item information, such as “if the recorded material is aluminium, assign the metal category”.
System B: uses a model trained on labelled images to predict the category of a new image.
Both use data and a method to produce a result. System A’s rule can stay fixed; System B’s model was adapted during training. Neither must alter itself every time a new item is processed.
The important distinction is how the method was obtained and whether it changes, not simply whether the output varies.
12 | COMMON MISTAKES
- Collecting data automatically means AI: explain how it is used.
- Reasoning proves human thought: the behaviour is simulated computationally.
- Every AI learns: fixed expert-system rules can be sufficient.
- A different output proves learning: a fixed method can respond to different inputs.
- Training and inference are identical: model adaptation differs from using it.
- More data guarantees better results: relevance and representation matter.
- Learning removes bias or errors automatically: evaluation is still needed.