By Jen Schellinck and John Stroud
It’s possible to categorize artificial intelligence (AI) in a number of different ways – e.g. based on functionality, techniques used, etc. In this short primer, a number of key terms and concepts will be put into context so that these categories and how they factor into the appropriate applications and use cases of AI can be more clearly appreciated.
Good Old Fashioned AI
Historically speaking, the first type of artificial intelligence researchers worked on is what is now referred to as ‘Good Old Fashioned AI’ (GOFAI). In the case of GOFAI, which is still an active area of research, intelligent behaviors are programmed by hand into a computer system using a series of rules and algorithms.
Under the right circumstances, these rules can result in useful and even quite sophisticated behaviors. A good example of this type of AI might be a basic floor-cleaning robot (e.g. a ROOMBA). It will have some movement rules hand coded in to its processor and in an appropriate environment the right simple rules will result in it relatively autonomously and intelligently cleaning the floors while at the same time keeping itself charged.
Expert systems are another type of GOFAI, which encode expert knowledge in what are called ‘decision trees’ and enable some types of decisions to be intelligently automated.
As AI research progressed, people realized that one strategy for generating artificial intelligence was not to program that intelligence into the agent directly, but rather to provide intelligent agents with the ability to learn from data they collected from their surroundings. This was accomplished by allowing them to alter their own programming (and by extension their behaviour) based on the collected data. This type of technique was broadly referred to as ‘machine learning‘.
To make machine learning possible, researchers developed a strategy with three elements: (1) learning programs, (2) rule templates (often referred to as models) and (3) rule running programs. These components would be, and still are, hand coded into agents.
However, the initially provided learning programs were then given the ability to alter the rule templates based on the data received by the agent. This altering of the rule templates would result in the creation of specific rules that were more appropriate for a given environment, based on the data provided about that environment. These specific rules would then be interpreted by the rule running programs under the relevant circumstances to dictate useful behavior within these ‘artificially intelligent agents’, where the agent might be either a computer agent (e.g. web-bots) or a physical agent (e.g. robots).
Types of Thinking Activities
In addition to these general approaches to artificial intelligence, researchers realized that artificially intelligent agents would need be able to carry out certain specific types of thinking activities. For example artificial agents would need to be able to anticipate what would happen (predict), create plans to achieve goals and carry out these plans, and make generalizations and abstractions about the world, among other abilities. Research was subsequently devoted to understanding how the two main strategies researchers had developed – GOFAI and machine learning – could accomplish these specific intelligence-related tasks.
As they worked towards these objectives, AI researchers also came to realize that many techniques that had already been developed by statisticians in order to help humans use data were equally applicable to artificial intelligence tasks. As a result these techniques – now often called ‘statistical learning’ – were added in to the tools available for creating intelligent agents. At the same time, statisticians and business intelligence analysts realized that techniques developed within AI research could themselves be borrowed to help not just artificial agents, but also people to better understand and use their data. Coming out of this cross-pollination, ‘data science’ and ‘data analytics’ have become umbrella terms encompassing techniques from all of these disciplines, particularly when these techniques are used in applied contexts.
GOFAI techniques, because of their full reliance on hand-coding, often fall into another AI category – ‘explainable AI’. Because GOFAI solutions are fully hand-coded, it is relatively possible to look at the code, follow its logic, and understand why an artificial agent is behaving in a particular manner under a particular circumstance. It is also likely that considerable understanding of the mechanics of a system or situation must first be made available in order to provide the GOFAI with abilities like prediction and generalization.
Somewhat surprisingly, machine learning is often not like this, in two important ways. First, because machine learning techniques effectively allow programs to alter themselves, often in quite sophisticated ways, it is possible for the program to achieve a particular result in a manner which is difficult to reverse engineer, once the rules template, or model, has been fully altered and optimized relative to the data provided. Thus machine learning it is typically not explainable AI, but rather ‘black box’.
Second, and related to this, the intermediate steps that humans would typically take in order to be able to understand and predict a system or situation – e.g. first learning about the mechanics and relationships between the different parts of the system or situation, then understanding how they lead to a particular outcome – are also not made visible or explicit in the case of machine learning results.
This is possible in the same way that it’s possible, with some practice, for people to use large amounts of incoming visual data to rapidly and successfully hit a baseball without actually being able to explain either how they have done so or which laws of physics had to come in to play in order for them to do so. Similarly, often a machine learning algorithm can predict an outcome without at any point needing to understand (or have us understand) in an explicit sense the components, elements or mechanics of the situation or how the AI itself came to make that prediction.
AI is a very quickly growing field, both in terms of basic research and applied results. There are many more details that could be discussed here, and new types and categories of AI continue to emerge. However, the terms discussed in this overview will provide the foundations for discussion of AI projects and support additional learning about this increasingly relevant technology.