Can artificial intelligence be used to create artificial intelligence? This was the question posed at the beginning of the recent Expert Lecture delivered by Professor Nelishia Pillay of the Department of Computer Science in the Faculty of Engineering, Built Environment and Information Technology (EBIT) at the University of Pretoria (UP). UP’s Expert Lecture Series provides a public platform for researchers to engage with a general audience on significant developments in their fields of expertise that are likely to have an impact in the future.

Using the example of finding the best route to a destination, Prof Pillay took the audience through the process of understanding how artificial intelligence (AI) systems work, and how they have developed from basic systems to knowledge-based systems, to the current era of machine learning, which is about getting AI to learn.

“There are three main areas of machine learning, the first of which is reinforcement learning,” Prof Pillay explained. “As I’m speaking to you, for example, I am gauging your responses; if it’s looking like what I’m saying is not being well received, I will change what I’m saying. In the same way, reinforcement learning takes an environment, throws actors in it, and then your agents decide what to do next based on what response is given. This is a popular method and is not computationally expensive.

“The next type of learning is called supervised learning, which basically learns the way we learn by making mistakes, then learning from them. Many people have heard of deep learning, which makes use of supervised learning through neural networks and is the most popular type of learning. It is expensive in that you need those experiences; you’ve got to get that data from somewhere. The last type of learning, unsupervised learning, has the advantage of not needing annotated data, and a lot of the research we’re doing is working towards unsupervised learning for that very reason.”

As the Multichoice Joint Chair in Machine Learning and the SARChI Chair in Artificial Intelligence for Sustainable Development, Prof Pillay then spoke about some of the interesting projects that she and her team have been working on that are related to innovation in industry, health and well-being, and education.

Prof Wynand Steyn, Dean of the Faculty of Engineering, Built Environment & IT

“With Multichoice as my sponsor, we do a lot of work in AI that is linked to broadcasting; one of the projects we’ve enjoyed working on is using AI for automated thumbnail selection for movies, using a combination of optimisation and machine learning,” she said. “When it comes to health and well-being, most of the work we’ve done is in disease diagnosis. Working with a chemical pathologist at UP, we automated the process for myeloma detection, which was straightforward machine learning – supervised learning. We also worked with UP’s Dentistry Department to detect oral lesions. Here we compared, supervised and unsupervised learning to see how that would work, and we got very good accuracies. Additionally, we’ve done a lot of work with cancer detection, and played with how AI techniques could work across different types of the disease, including skin cancer, brain tumours and lung cancer. We’ve also used optimisation and machine learning for the detection of heart disease, COVID-19, diabetes and depression.”

AI developments for education is where Prof Pillay’s passion lies, specifically those related to lifelong learning.

“Graduating from university is just the start of learning, because the technologies we use and what we do in our careers are going to change,” she explained. “We need tools to support lifelong learning, and AI can be used to do this, by first identifying the skills we’re going to lose, then by assisting with the relevant training.”

Prof Pillay said her journey with AI started with intelligent tutoring systems. One of her projects involved rolling out an intelligent tutoring system that could be used for speech therapists in South Africa.

The lecture worked its way to answering the key question posed at the beginning, and which is on the tip of everyone’s tongue: Are we at the point where AI can be used to create AI?

“To solve a problem, you need some knowledge of neural networks,” Prof Pillay said. “It’s laborious for computer scientists to tune the necessary parameters; that’s why we’ve asked the big question – can we get AI to do this for us? There’s been a lot of work in an area called automated machine learning, and the short answer is that we’re at the point where AI is creating itself. The challenge is to get a non-expert to use it, so we’re looking at how it can be packaged so that you can go out and buy the software.”

Doesn’t this open AI up to be used for nefarious purposes?

“At a recent conference, this same question was asked; and the general response was, ‘We don’t need AI to be unethical; we are capable of being unethical without it,” said Prof Wynand Steyn, Dean of the EBIT Faculty, adding that just as we are capable of working with good intentions, so too can AI be used to support such endeavours.