Artificial intelligence (AI, also machine intelligence, MI) examines how we can achieve intelligent behaviour through computation, i.e., with programs or machines that can mimic human intelligence, i.e., they are adaptive, effective, and safe in a variety of new situations.
AI is an interdisciplinary field: it draws from cognitive science, mathematics and computer science, philosophy (in logic), engineering (in control theory), economics, psychology, and neuroscience.
Basics
The primary goal of AI is to maximise the expected utility of actions, i.e., to make rational decisions in a search space where we don’t have full knowledge or control, and also where we don’t know precisely what effects our actions might have. An agent is an entity that perceives the environment, and acts accordingly (much like a system controller in control theory).
There are two primary approaches to AI.
- Symbolic AI uses explicit symbols and rules to represent knowledge and perform reasoning. This allows agents to logically conclude how to perform a task without needing a large dataset, and is suitable for tasks requiring clear logic and reasoning (like automated theorem provers).
- Sub-symbolic AI learns patterns from data, i.e., traditional machine learning or deep learning. These are popular because they scale well with large datasets and are suitable for tasks involving perception and pattern recognition.
Statistical and symbolic approaches can be combined. For example:
- We can embed a symbolic reasoning engine inside a neural engine, like ChatGPT using the Wolfram mathematics library.
Key concepts
- Symbolic AI
- Sub-symbolic AI
- Generative AI
- Sub-fields
- Philosophical ideas
Resources
- Artificial Intelligence: A Modern Approach, by Stuart J. Russell and Peter Norvig (AIMA)