When computers display the type of intelligence that humans and animals possess, this is known as artificial intelligence (AI). Intelligence involves perceiving the environment and using this information to take actions that maximize the likelihood of achieving objectives. Specific AI capabilities include learning, perception, reasoning, planning, knowledge representation and natural language processing.
Ultimately, it is hoped these can be combined into a ‘general intelligence.’ These functions contrast with the clearly defined algorithms of non-AI computing. AI is a general term for any artificial entity that has these abilities, but in practice, it is generally realized using a neural network to perform machine learning.
Artificial general intelligence (AGI) is the ultimate aim for many AI researchers. It would enable a machine to understand or learn anything that a human can, without requiring an pre-programming or being pre-configured by a human operator. AGI is sometimes referred to as ‘strong AI,’ in contrast with the ‘weak AI’ or ‘narrow AI.’ Weak AI is currently used to perform a very specific task for which it has been given clear success criteria and training data.
AI should not be confused with self-awareness, or artificial consciousness. AI can already perform certain narrow tasks with a far higher capability than a human, without any reason to believe it has any self-awareness. It therefore seems possible that a general intelligence would not require self-awareness. Understanding the nature of self-awareness is highly philosophical and we are far from reaching any consensus on how it arises in humans, never mind how it might be artificially created.
Artificial intelligence is sometimes regarded as a catch-all term for tasks not yet, or only recently, resolved by computing technology. Therefore, functions such as optical character recognition, which was once regarded as AI, are no longer considered to be ‘intelligent.’ Similar capabilities, which remain at the cutting edge of computing, are still considered to be examples of AI — for example, speech recognition and autonomous vehicles.
Most current developments in AI use artificial neural networks, networks of connections that can store and process information. These are well-suited to learning, requiring only an objective, with no need for precise instructions on how to perform the task. The way that neural networks learn, using mathematical models, is closely related to regression in non-linear statistics. However, while regression might be used to fit a line to data with two dimensions, neural networks fit models to data with much larger numbers of dimensions. This means they require very large datasets, although it is interesting to note that human intelligence can sometimes learn much more quickly with little data.