Computers are increasingly able to accomplish tasks that are difficult for human experts, such as diagnosing diseases or detecting credit card fraud. While the earliest examples of computational thinking can be traced back to the 13th century, according to Holger Hoos, Leiden Professor of Machine Learning, we are now entering an age of advanced computation, in which it is becoming increasingly important to balance the risks and rewards of artificial intelligence (AI). Inaugural lecture 27 October.
Extend rather than replace
It is important to make sure that artificial intelligence extends rather than replaces human intelligence, Hoos believes. In his inaugural lecture, he sketches the history of computer science, linking antique mechanisms for navigation and astronomical predictions with the work of medieval philosopher Raymundus Lullius and Charles Babbage’s design of the first mechanical computer at the height of the industrial revolution. Computers were originally conceived to perform complex numerical computations faster and more precisely than skilled human calculators. Yet, Hoos argues, highlighting the vision of 19th-century mathematician Ada Lovelace, it has long been known that computers and the software running on them could do much more.
Automating the programming of computers
Algorithms are at the heart of programming computers to perform all sorts of tasks, from surfing the web to diagnosing cancer. They are like extremely precise cookery recipes in that they specify precisely the steps that need to be carried out to process given data. Algorithms are, Hoos says, precise instructions that, flawlessly executed, produce a specific result or behaviour. Finding good algorithms and making them run efficiently on computers is difficult, which is why learning to programme computers is something many people find challenging. It requires a specific mind-set, Hoos believes, and even then it can be surprisingly hard to do.
Deep neural netwerks
Machine learning is essentially about automating the programming of computers, about finding good algorithms automatically. There are different approaches to this, Hoos explains. A particularly successful one, at least in situations where lots of data is available, is inspired by the physiology of our brains. These so-called deep neural networks are instrumental in the development of self-driving car technology, as well as for many tasks in image and language processing, such as face recognition and automatic translation. Still, Hoos cautions, deep learning is only one of several approaches, and it has its limitations; for example, it is very difficult to understand the limitations and hidden biases of a deep neural network – an important topic for research. ‘There are also some challenges, such as looking for techniques that can learn from fewer data, using less powerful computers, in a way humans can understand. That’s something I want to focus on,’ Hoos says.
Programming by Optimisation
At the technical level, most work in machine learning deals with a few very specific problems. Hoos pursues a more general approach that applies learning to better solve a large range of problems that are challenging for humans, using advanced optimisation techniques to find among the astronomical number of programmes for a given task ones that work particularly well. This, according to Hoos, who pioneered this paradigm known as programming by optimisation, ‘changes how we programme computers in a way that combines human creativity and ingenuity with algorithmic efficiency and lots of computation.’
Hoos states that computer science has two historical roots, in mathematics and engineering, plus a third foundation that has been added more recently: empirical science. As computer hardware and software has reached a level of complexity similar to that of biological organisms, controlled experiments and statistical analysis have become indispensable tools for computer scientists. Hoos: ‘Empirical methods are particularly indispensable in artificial intelligence and machine learning, where progress critically depends on the use of heuristics -‘rules of thumb’ – whose efficiency is beyond mathematical analysis.’ Many of the most challenging problems in computer science, Hoos explains, resemble the search for a needle in a haystack, and good heuristics serve as shortcuts in this kind of search.
The dream of machines that match human intelligence across the full range, from logical thinking to arts and science, is quite old, but there is an increasing feeling that it can be made reality, perhaps not too long from now. ‘No one can predict how long it’s going to take to get there,’ Hoos explains, ‘but once machines reach full human-level intelligence, they’ll likely surpass our capabilities very quickly and across the board. There is no reason for any form of panic, but now is the right time to think seriously about the consequences, the risks and rewards.’
Human level of artificial intelligence not desirable
Hoos emphasises that these discussions should involve not only computer scientists, but also academics from other disciplines, politicians and the general public. ‘I’ve come to the realisation, he says, that general human-level AI, while intellectually fascinating, is not all that desirable. Instead, we should focus on AI that complements our abilities and compensates for weaknesses. Evolution has equipped us for living in a world that is primarily determined by local phenomena and short-term interactions. But evolution is much too slow to adapt us to the long-term, global effects of our actions and decisions. We should focus on developing AI that helps us overcome our limitations, responsibly manage our limited resources, and interact with each other constructively, respectfully and enjoyably. We need AI that extends, rather than replaces, human intelligence.’