Machine-learning is a boon, but it still needs a human hand
We should remember that devices only compute what has been programmed by people
Advances in computer power, machine-learning and predictive algorithms are creating paradigm shifts in many industries. For example, when an algorithm outperformed six radiologists in reading mammograms and accurately diagnosing breast cancer, this raised questions around the role of machine-learning in medicine and whether it will replace, or enhance, the work being done by doctors.
Similarly, when Google’s AI software AlphaGo beat the world’s top Go master in what is described as humankind’s most complicated board game, The New York Times declared “it isn’t looking good for humanity” when an algorithm can outperform a human in a highly complex task.
Both these examples point to narrow uses of artificial intelligence, specific types of machine-learning that are hugely effective. The medical example illustrates supervised learning, where a computer is programmed to solve a particular problem by looking for patterns. It is given labelled data sets, in this case X-rays with the diagnosis of presence or absence of breast cancer. When given a new X-ray, the computer applies an algorithm based on what it has learnt from all the previous X-rays to make a diagnosis. Unsupervised learning is a sort of self-optimisation where a computer has a set of rules, such as how to play Go, and through playing millions of games learns how to apply these rules and improve.
What is machine-learning?
Machine-learning is a phenomenal tool. To fully harness its potential it is essential to understand what machine-learning is (and isn’t) and to demystify some of the hype and the fear around what it can and can’t be used for. We have anthropomorphised computers; we speak about them in terms of intelligence and learning. But in essence, a machine computes — it does not learn. Its algorithms are designed to mimic learning. In essence, these algorithms minimise the errors of a complicated function that maps inputs to outcomes and we interpret that as solving a problem, but the machine doesn’t know what problem it is solving or that it is playing a game. The intelligence rests with the humans who design the algorithms and configure them for specific tasks.
Now, more than ever, we need intelligent and well-educated people who can apply these techniques in the correct context and interpret the results. When an algorithm fails, the consequences can be catastrophic. An obvious example is a fatal accident caused by a self-driving car. We need to build in fault tolerance. Data integrity is also an important issue — what we put in is going to affect what we get out. Education is critical in making sure we get these elements right. And, of course, there are broader ethical issues to consider surrounding data collection, such as what data can be used, where it is sourced, and whether different data sets can be combined.
Machine-learning is particularly valuable in the financial sector. Many applications are already in use in banking, insurance and asset management. Financial institutions use pattern recognition successfully for fraud detection. It is also valuable for looking at trends in data sets and finding patterns that humans may not be able to identify directly, for example in profiling people who apply for credit. There are even robo-advisory applications for individual asset allocation. In financial modelling, machine-learning can be applied to pricing, calibration and hedging.
For example, valuing derivatives contracts depends on many complex factors and variables such as interest rates, exchange rates, equity values — all of which fluctuate all the time. Financial mathematicians use models for this, but they are complicated and not easy to solve in a closed form. We may be able to build and apply a model to one contract, but banks have hundreds of contracts, and risk management and regulatory frameworks need to be updated all the time. Machine-learning, specifically deep learning and neural nets, provides a powerful shortcut. We can use classical numerical methods to produce financial models and then use them as labelled data sets — as in the X-ray example. An algorithm can take this input to generate the output for multiple contracts.
Industries and organisations that are pulling ahead are figuring out where to replace standard methods and complex, time-consuming computations with machine-learning. They are also using it for more complex modelling approaches, adding further variables that cannot usually be factored into standard methodologies. The most obvious benefit is that it is faster — machines can compute millions of times faster than humans. These techniques also have the potential to be far more accurate and allow us to make better-informed decisions.
But the human element is critical. The accuracy of potentially life-changing outcomes will depend on how we identify where we use these techniques, how we build the algorithms, how we choose and manage data and, finally, in how we interpret and act upon the results.
• Prof McWalter is an applied mathematician who lectures computational finance at UCT’s African Institute of Financial Markets and Risk Management. Prof Kienitz lectures at the University of Wuppertal and is an adjunct associate professor at UCT. His research interests include numerical methods in finance and machine-learning applied to financial problems and derivative instruments.