So what is artificial intelligence (AI)?
The goal of AI is largely to develop hardware and software that is capable of doing things which are normally done by people, and in particular, done by people acting intelligently1)http://www.computerworld.com/article/2906336/emerging-technology/what-is-artificial-intelligence.html. Some of the most recent well know examples are programs which beat human experts at Chess, Go, Poker etc.
Over the last 60 years, the goal posts of AI have moved a significantly.
One of the early successes of AI was to read handwritten post-codes of letters for the postal service.
At the time, handwriting recognition was cutting edge AI, today there are numerous apps on smart phones which can do this.
Similarly, face recognition when first developed was a breakthrough, now the most basic digital camera offers the functionality.
These examples of AI are typically grouped as narrow or weak AI. Narrow AI is focused on one narrow task and can't be generalised to other tasks and isn't sentient.
The AI often hyped in media is general or strong AI 2)https://en.wikipedia.org/wiki/Artificial_general_intelligence. General AI, often referred to as Artificial general intelligence, it would be sentient, but it doesn't exist yet. A general AI solution would be able to perform any intellectual task that a human can.
While no general AI exists yet, there have been significant advances in AI over the last five years in particular (since 2012), as various flavors of neural networks etc have been developed, and as compute power has continued to increase.
An often used test for AI is its ability to identify what is in an image. Its trivial for humans to identify objects in pictures (e.g. a tractor), but when you consider all the different angles, shapes, colors, lighting, environment etc possible, it's a hard problem for a computer to manage. Or at least it use to be.
The use of neural networks (in the form of deep learning) has made a big impact in image recognition since 2012, and is now more accurate than humans.
Neural networks are one of the main methods for AI discussed in the media, and deep learning is, simplistically, a very complex neural network.
Neural networks basically mimic the way humans learn. They are trained by providing them with a large amount of labeled data for the problem in question (e.g. image recognition), and adjusting weights within the network so that the highest probability output matches the label. Its synonymous to teaching a child to read by them multiple repeated examples. Give a neural network a large enough data set, and it can generalise to the correct output.
But neural networks currently need very large data sets (millions or billions of labeled examples) in order to generalise, and a lot of computing power to train, before they can be used to identify inputs (e.g. images) they haven't seen before. Humans are far more efficient learners. The big change in the ability of neural networks over the last five years has largely been due to amount of labeled data available (e.g. from the internet), and a combination of rapidly increasing compute power and improved training algorithms.
AI is currently capable of, or being developed to, answer general questions (e.g. IBM Watson), translate text between languages (e.g. Google translate), read an interpret text, drive cars etc.
References [ + ]