The families of AI: symbolic and learning
Two ways to build an AI, writing the rules or having them learned from examples: expert system, inference engine, what each family can explain and demands, and the key dates from 1950 to 2022.
- 24 cards
- 5 topics
- Free
The outline, topic by topic
Each topic is weighted the way the exam weighs it.
- Concepts
- Questions
- Sort and link
- On the timeline
- Understand
A few cards from the deck
Symbolic AI: what is it?
Show the answer
The family of approaches in which humans explicitly write the knowledge and the reasoning rules, which the machine then manipulates.
Every conclusion can be traced rule by rule, but everything the machine knows had to be written by hand.
A forward-chaining engine starts from the facts “fever” and “cough”. Which facts does the base hold at saturation?
```text
R1: fever AND cough -> infection
R2: infection AND tiredness -> flu
R3: fever AND rash -> measles
R4: infection -> rest advised`- Fever, cough, infection and rest advised
- Fever, cough, infection, flu and rest advised
- Fever, cough and infection, because R4 has no second condition
- Fever, cough, infection, measles and rest advised
Show the answer
Fever, cough, infection and rest advised
R1 fires and adds “infection”. That makes R4 fireable: “rest advised” is added. R2 needs “tiredness” and R3 needs “rash”, both absent: they do not fire.
In …, the engine starts from the known facts and fires every rule whose conditions are met, until …: no new fact left.
Show the answer
In forward chaining, the engine starts from the known facts and fires every rule whose conditions are met, until saturation: no new fact left.
A fact that was added can make another rule fireable: that is what moves the chain of deductions forward.
Machine learning: what is it?
Show the answer
The family of approaches in which the machine adjusts its own parameters from examples, instead of following rules written one by one.
The behaviour comes from the data: better examples improve the system, bad examples degrade it.
Why is a spam filter usually learned rather than written as rules?
- Senders keep changing their wording: retraining on new examples is easier than rewriting rules
- A written rule cannot test whether a word appears in a message
- A learned model justifies each rejected message better than a rule does
- Rules need more example data than a learned model does
Show the answer
Senders keep changing their wording: retraining on new examples is easier than rewriting rules
Rules are readable but quickly bypassed; learning follows how messages evolve. It is, however, less explainable and depends on the quality of the examples.