I am writing this to clear up a lot of misconceptions that I read about AI, not to give _a_ definition, but rather to give _some_ definition to it.
So first, a first attempt. Artificial Intelligence is a computer method (rarely mechanical) that seems to do something only a human could do. A talking responding machine is AI, but a walking machine, with articulated legs, is not. Unless... well there it is, sometimes it is the application that is AI and sometimes it is the methods. And a chess playing machine....well is that AI or just really good engineering, exploring all possible game possibilities quickly?
We already have two nuances: how an AI is engineered, and what does the label apply to. The methods of AI engineering fall into two broad areas: human simulation, attempting to mimic the internal biological/psychological processes, and direct engineering, doing what it takes no matter what to make the engineered device act externally intelligent. A language parser uses scientific theories of linguistics to mimic how the brain is supposed to manipulate the text of a sentence to determine understanding. A chess playing machine uses alpha-beta pruning of deep game trees, contrary to the usual method people which is to look at only two or three moves deep and vaguely judge 'strength' of a position based on experience.
A lot of what was once considered AI is now considered plain old engineering. Once you see under the covers that the method used to look outwardly human is only a boring step by step recipe or lookup table, it loses it's 'wow' AI appeal and just seems like regular non-AI computing. in the 1600's, calculating machines were magical and might have been considered AI (if that were a thing then) because most people thought that people themselves doing it was magical much less a machine (OK this is a bit hyperbolic. what is 'most'? who are these people?). But you get the idea. A machine doing chess is magical (or has a small hidden turkish chess master manipulating the controls within). But Deep Blue in 1995, which beat Kasparov the chess grand-master, was pretty simple just using search trees, no ethereal silicon-embodied intuition ('just' = very rocket-sciency search trees and some hand-curated openings, gambits and end games).
An AI, with the indefinite article, is by popular usage, the thinking talking machine. Because of the potential for confusion the term for that nowadays is usual referred to as an AGI, Artificial General Intelligence.
But now to more substance. All along we were going along in vague style, not really knowing what AI really refers to concretely, relying on a presumed common idea of AI without really knowing. So here are some concrete examples of what AI is.
First, applications, which is what externally looks like everybody's general idea of AI
- vision - converting digital images into words
- language - speech recognition, natural language processing and understanding, chatbots
- games - chess, go, crosswords
- problem solving/reasoning - word problems and puzzles, expert systems
- planning - taking a set of goals and initial conditions
- robotics - not the mechanical part of for example an articulated hand grasping an object but coordinating a path over a complicated landscape or separating parts on an assembly line
- learning and memory - most of computer science has been intent on solving the memory problem and has done it well. Financial statements, medical records, airline reservations, online libraries.
And then there's tools and methods, how it's actually done, the man behind the curtain.
- logic and other classical mathematical methods - for reasoning
- probability/statistics - most machine learning methods are within the realm of stats
- combinatorial algorithms - when a perfect algorithm for one domain works most of the time for another, then it's called a heuristic
- decision trees - chess? one big decision tree? Yes.
- optimization - specifically numerical algorithms for optimization like simplex or A*
- cognitive science - cognitive psychology, neurology (brain physiology, neurons), linguistics, so you know how it works in actual biology, so you can either be inspired by it (neural networks) or simulate it (parsing a natural language according to linguist's rules)
The man behind the curtain is an apt metaphor for AI because like in Oz, the apps look magical, but it turns out it is, well, not exactly charlatany; the methods are so...mechanical and spartan and inhuman and small tricks. For example, crosswords seems like you need to have not just broad and deep knowledge, but also clever ability with associative memory and puns to read behind the clues. The recent very successful AI methods for crosswords is just to use a dictionary search and almost ignore the clues themselves, by any means necessary no matter how cleverless.
With all that said, the motivation for this was what I find to be a lot of misuse of all these labels, one term being used for another. So here's a bullet point summary:
- AI is what you call the big show, all the different things together
- AGI (Artifical General Intelligence) is a very limited example of an AI app but the most salient, the 2001/HAL intelligent like spoken interface. This seems to be like the end goal of AI but currently unrealistic, or rather very realistic with very limited expectations.
- ML (Machine Learning), which is mostly just statistics, is a tool to quickly create AI apps from lots of data without having to rely on a domain expert
- ANN (Artificial Neural Nets) and DL (Deep Learning, which is just a big NN) is just (of many) ML techniques
No comments:
Post a Comment