What is Artificial Intelligence?
In 2018, Matt Velloso wrote in Twitter:
"Difference between machine learning and AI:
If it is written in Python, it's probably machine learning
If it is written in PowerPoint, it's probably AI"
The tweet has tens of thousand of likes. It is a joke, but it is a good basis for starting to think about ML (machine learning), AI (artificial intelligence) and the hype around those terms. If you do not understand the joke, don't worry. You will after reading this lesson material thoroughly.
In order to have a non-controversial definition of artificial intelligence, we would have to have a definition of intelligence itself. Sadly, we don't. For that and other reasons, the scientific community doesn't have a single, rock-solid, waterproof definition of AI.
The next two paragraphs are based on a book Artificial Intelligence: A Modern Approach (S. Russell & P. Norgiv. 3rd ed., 2010). According to the book, AI is a field that "attempts not just to understand but also to build intelligent entities." Definitions of AI can be organized into four categories: thinking humanly, thinking rationally, acting humanly and acting rationally. The split of these four categories is based on two axes: rational-humanly and thinking-acting. If an AI is acting humanly, its behavior is hard to differentiate of one with human. For example, an interrogator would not know if she is chatting with a bot or a human. An opposite to this on both axes is thinking rationally. A rational thinker bot in this case would have an ideal reasoning process. Everything would be based on fully flawless logic. Can you imagine what would be the obstacles predicting/solving real-world data and problems using this kind of bot?
Rational-Acting seems to be the pair that is most useful approach for practical AI. A rational agent is an entity that perceives its environment using various sensors and act accordingly, but is able to adapt to change and pursue goals. During this course, we will be creating a number of various kinds of rational agents and their components: machine learning is one and currently the dominant way of building an AI.
Branches of AI
AI is an umbrella term consisting various of branches or fields. In a book Artificial Intelligence with Python by Prateek Joshi (2017), the dominant topics are listed as:
- Machine learning and pattern recognition: learning from data. Most of the AI performed within this course fits into this branch.
- Logic-based AI: a set of statements that can be written in the form of a decision tree, and is often used in language parsing.
- Search: these programs find a path within a network. Used in strategy games and path finding in navigation applications.
- Knowledge representation: defining relationships within elements in a domain, e.g. in a form of a taxonomy or other hierarchical structure.
- Planning: planning based on known facts of the world.
- Heuristics: finding paths or solutions in a situation where optimal solution may not exist or is non-practical to solve.
- Genetic programming: programs that are mating agents and selecting the fittest.
It is not required or even practical to try to map and memorize all branches of AI and their definitions. Key thing to remember is that AI includes a wide range of approaches. It comes in many forms and shapes, and appears in many different kind of applications such as in computer vision, natural language processing and robotics.
The history and the hype
Artificial intelligence isn't a new invention. Trying to understand reasoning of a human mind, or to construct a mechanical device that can do some form of calculations, has been going on for hundreds or even thousands of years. The first neural network computer, SNARC, was built in 1950 by Minsky and Edmonds. It consisted of 40 artificial neurons, which are inspired of the functionality of neurons in human brain. In comparison, a fruit fly has around 100 thousand neurons. In 1950s and 1960s, computers were already solving algebra story problems and playing checkers. These years were full of hope and great dreams. AI researchers made predictions and promises about their future success stories: a computer would soon win a human in chess - maybe in 10 years. From mid-1960s until early 2000s was a period of time that had many ups and downs, or "hype cycles". In midst of this era, in 1980s, there was a strong but brief moment of hype and high hopes. Billions of dollars were spent on various applications of AI, followed by disappointment. This started another downfall, or "AI Winter". Currently, we are living at a new arrival of AI which is often called with a name "AI spring". Hopes and expectations are high, again. We have more data and computing power than ever and AI has been used for business since 1980s, so there surely are reasons for hype. Will we have a new AI winter? No one knows for sure. Usually, predictions about soon-to-be technology tend not to be too trustworthy.
Above is a graph from the 2019 AI Index Report by Stanford University. The interest in AI has fairly suddenly increased. It is highly recommended to check the AI Index Report and read at least the beginning of each chapter.
Task
Using your favorite search engine, find answers to these questions:
- What is the Turing test?
- Who or what is Eugene Goostman?
- What are Narrow AI and General AI?
- What is Gartner hype cycle and what are the 5 stages of it?
- Can one find a definite list of all of the branches of AI?
Write the answers to your notebook (either digital or analog.) These topics may be covered in the exam. You should be using at least 2 hours researching the topic and reading the articles or watching videos.