Introduction
Contradictory definitions of AI
Defining AI through
- thought processes, reasoning vs. behaviour
- similarity to humans vs. rationality
Thinking humanly
We need an understanding of the internal activities in the brain (thoughts).
There are 2 approaches to check our progress:
top-down Cognitive Science - predicting and testing behavior of human subjects
bottom-up Cognitive Neuroscience - Gathering neurological data
Problem: We dont know what thinking humanly is - available theories cant describe human-level general intelligence.
Acting humanly
Human abilities:
choosing suitable abstraction level, commonsense, natural language, (real world)-reasoning.
Turing Test / Imitation Game:
Test for intelligent behaviour - how long can a lay person be fooled thinking its talking to a machine?
Not relevant anymore. Can not be formalized.
Thinking rationally (Lawful thinking)
- Logic: What are the rules of thought?
- Philosophy: idea of mechanization of thought (letting a machine decide whether a thought process was correct)
Problem: intelligent systems aren't always rational
Acting rationally
Rational behavior = maximizing uility based on available data (Thinking not required)
Problem: computational limitations make perfect rationality unachievable
Shortcomings
Currently AI is is not capable of:
- Abstraction
- Natural Language Understanding
- Question Answering
Technological perspective
- scalability and efficiency
- robustness
- validation and verifiability
- explainability / traceability — we want to know the line of thought that led to a specific result and want to verify the solution path