What does a machine need to know to understand language? In this video, learn about finding words and their relationships.
- Symbolic reasoning and planning AI worked great for machines whose matching patterns were limitive. You could easily use this approach for creating a program to help you with your taxes. The IRS only has a limited number of rules, and you can go through the paperwork step by step. You could even create many if-then statements. If you have a dependent, then fill out x form. If you live in this state, then fill out the z form. The challenge with this approach is that if the list of possibilities gets too long, then the system is difficult to manage.
Let's say you wanted to create an AI program that identified animals from a database of images. You wouldn't want to create a list of these matching patterns, that would have too many variables. Imagine all the different lists of size, eye shapes, number of legs, or other distinguishing features. You also wouldn't want to create something like if it has fur, then check to see if it has whiskers, if it has whiskers, then check the shape of its ears. A list like this would become way too long if you wanted to match many different kinds of animals.
There are also way too many opportunities where the system could get stuck. Maybe you couldn't see the whiskers in a low resolution photograph. Then you have an unidentified image. That's why early AI researchers started to wonder if instead of planning out matching patterns, a computer could be programmed to learn the new patterns by itself. They call this idea machine learning. Machine learning got its start very shortly after the first AI conference.
In 1959, an AI researcher named Arthur Samuel created a program that could play checkers. This program was different from any of the other symbolic systems. It was designed to play against itself so it could learn how to improve. After each game, it would learn new strategies, and after a short period of time, it consistently beat its own programmer. This was a breakout discovery. For one, you didn't need an expert to create symbolic patterns.
So it wasn't like an expert system that needed a doctor to list out all the possible responses. The machine would create the expert list and also look for the matching patterns on its own. Think of it this way. The Chinese Room Argument had the computer match the incoming notes with a book of Chinese phrases. Then the computer looked through the phrases and taped together a response from the Chinese characters on the floor. That's a tedious job, yet the really tedious job was creating the phrase book.
Imagine having an expert Chinese speaker guessing through all the notes that might come through the slot. Instead you could try to use machine learning. Here you have a program watching notes being passed between two native speakers, almost like it was processing email conversations going between two people. After looking at many thousands or even many millions of conversations, the machine could start to pick up patterns. It might notice that conversations always start with similar greetings and responses, so it would create its own book based on these matching patterns.
Something like if the note says this, then be sure to say this as a response. You still run into the same challenge that John Straw pointed out with the symbolic systems approach. The machine doesn't exactly know what's being said, it just matches the symbols and identifies patterns. The big difference then is instead of having an expert speaker list every phrase, you have the machine learning the patterns by looking at the data. One of the big advantages to this approach is that the machine can continue to learn.
If the machine finds out new patterns, it can rewrite the book to make it more accurate. With planning AI and the symbolic systems approach, you're stuck with the book that was written by an expert. In the past few years, machine learning has been one of the fastest growing areas in AI. That's partly because it's now much cheaper to store massive amounts of data. In fact, sometimes machine learning is used interchangeably with the terms data science and big data. If you think about this, it makes a lot of sense.
Organizations are collecting vast new amounts of data. So now, the big challenge is to figure out what to do with it. Machine learning can help you identify patterns without even really knowing what you're looking for. In a sense, you have artificial intelligent machines learning what's in your data and letting your organization know what it has found.
This course will introduce you to some of the key concepts behind artificial intelligence, including the differences between "strong" and "weak" AI. You'll see how AI has created questions around what it means to be intelligent and how much trust we should put in machines. Instructor Doug Rose explains the different approaches to AI, including machine learning and deep learning, and the practical uses for new AI-enhanced technologies. Plus, learn how to integrate AI with other technology, such as big data, and avoid some common pitfalls associated with programming AI.
- The history of AI
- Machine learning
- Technical approaches to AI
- AI in robotics
- Integrating AI with big data
- Avoiding pitfalls