AI and machine learning

The first in a series of blogs on AI by Dr Jeroen Vendrig, from Canon Information Systems Research Australia (CISRA)

The term AI may conjure images from science fiction movies where almighty machines mirror or exceed human capabilities. Such systems are known as Artificial General Intelligence (AGI), which are a step up from the so-called “weak” or “narrow” AI algorithms entering our daily life now. Don’t let the name fool you, weak AI is very powerful in practice. Unlike AGI, AI is very good in a specific task only, e.g. determining chess moves to beat the human champion. But that same AI can’t do other tasks that seem trivial to humans, such as physically moving chess pieces between squares without knocking over the other pieces. Several AI modules can be integrated into a full system. But still, that system can perform the predefined task only. It may beat you in chess, but it can’t play Snakes & Ladders.

AI has been around for more than half a century, starting with expert systems encoding human rules and fuzzy logic to control devices such as washing machines. Over the years, AI has evolved from having to receive detailed instructions to learning by example, closer to how humans learn. Advances in the field of machine learning have triggered an avalanche of AI applications that make machines smart rather than just work horses.

Weak AI may just be a one-trick pony, but it can do that trick very well and tirelessly. It’s up to us humans to select and design the right AI for the task. So what are the options? In this post, we explain the general types of tasks you can select to drive machine learning. In a later post, we’ll some in to more specific tasks for imaging and computer vision.

Task type	Description	Examples
Classification	Given historical data for an individual, provide a predefined category.	Given her history, will this student pass the test? Given his profile, which product will a customer buy?
Regression	Given historical data for an individual, estimate a value for a specified measure.	Given her history, what mark will this student achieve in the test? Given the business type, how long will it take this customer to pay?
Recommendation	Given historical data for an individual and multiple options, provide the preferred option.	Given his interests, what topic should the student study? Given previous selections, what’s the best photo to print for this person?
Clustering	Given historical data for a group, find sub-groups with similar characteristics.	Given their histories, what types of students are in a class? Given purchasing details, what are the customer segments?
Association/discovery	Given historical data for a group, find relationships for further analysis.	Given historical results, high marks are correlated to students who use blue inked pens.

Selecting the right task type is an important first step for deploying an AI system. Often, it makes sense to redefine a problem. For example, you may start with wanting an estimate for a student’s mark (a regression task), but end up with a classification problem instead: will the student get a low/medium/high mark.

In the above task type descriptions, the focus is on the final output from an AI system. Once that is clear, the question begs: how is the AI going to learn to produce that output? In the next two posts, we’ll discuss the learning material and how an AI knows it’s on the right track.