At least 27 sessions at Sibos in Sydney covered the topic of AI – more than those looking at application programming interfaces (APIs) and definitely more than quantum computing.
As Dr Ayesha Khanna, chief executive of Singapore-based incubator and advisory firm Addo AI Services, told Sibos delegates: “AI is one of the core pillars of the fourth industrial revolution.” It would seem that AI is, at last, coming of age and as someone who studied and worked in AI in the lean years of the 1970s and 1980s, I find this particularly gratifying. But is this yet another false dawn?
Jörg Hessenmüller, head of group digital transformation at Commerzbank, told Club@ Sibos that AI will help the bank to make use of the huge amount of internal and external data it collects and analyses to provide clients with tailored offerings. “Already today machine learning plays an important part in our business,” he said.
The promise of a true mechanical intelligence is almost as old as computers themselves. Alan Turing predicted in 1950 that by the end of the century people would speak of computer “thinking” without fear of contradiction and that people would take computers for walks in the park and exchange witticisms with them. But we soon discovered that these wonderful new machines were excellent at summing columns of numbers but utterly floundered when trying to understand a simple typed English sentence. Computer chess – a popular challenge for the pioneers, many of whom played the game themselves – was, to put it kindly, very weak at best. Worse, it turned out that everyday activities that most human beings can do, such as visual processing and navigating around in the real world, were simply impossible for an automaton.
Then in the 1980s, along came expert systems. The expert system had limited scope. There was no robotics, vision, voice recognition or other fancies, just a generic application which could be trained by an expert (a doctor, lawyer or financial adviser, for example) to do a specific expert task. Once trained it would be available for all time, tireless, unerringly accurate and, of course, much cheaper to “run” than the human expert who trained it. Some of today’s machine learning systems have a similar approach. But in the end there were only a handful of successes, mainly because the experts found them too difficult to train. Basically their expert reasoning couldn’t be expressed using the limited ad hoc user interface provided by the expert system itself.
After these earlier failures it would seem that, at last, there is good reason today to be optimistic for the future of AI and, along with many other industries, it will have a considerable impact on banking.
First, there is Moore’s Law, which states that every 18 months computer power – both processor speed and memory capacity – doubles. That may not sound so spectacular but over the decades the cumulative effect is enormous. For example, the inexpensive and unexceptional notepad computer on which I’m typing this article has 2000 times as much memory as and 300 times the processor speed of the state of the art DEC PDP-10 which I used in time-sharing mode when a research student at Edinburgh University’s Department of Artificial Intelligence in the late 1970s. Imagine: 300+ DEC-10s, no longer housed in air-conditioned fortresses, but dumped in my day bag alongside my lunch box!
I’ll spare you a detailed comparison of the old DEC with a real research tool, such as IBM’s Summit super-computer, where the figures are truly eye-watering but, in brief, the Summit is billions of billions of times bigger and faster. And with that amount of grunt, even a stupidly naïve AI program can start to look somewhat smarter and maybe even achieve something practicable. We’ve seen in part what raw processor power can do for computer chess – from bumbling amateur in the 1970s to world championship level in 1996.
Second, we are now in a digital world. If the expert systems of the 1980s had all that data to draw upon, things might have been different but, at the time, there was nothing, only what was lodged in the expert’s brain. Nowadays practically all information is available in digital form – words, pictures, sounds and knowledge along with proprietary data held in state and company archives. For example, machine translation now works – albeit crudely at times – thanks to the vast body of (human) translated text available online. And it turns out that no great insight into human language processing is needed – in essence a translation system merely finds a close match of the supplied text with something in the immense archive of translated works created over the decades by human translators.
Finally, we have machine learning (ML) systems. ML, especially in the shape of artificial neural networks was once much discredited by mainstream AI researchers. Now, extraordinarily, the networks have been found to work and work very well. A recent spectacular breakthrough was a deep neural network learning system called Alpha-Go, which beat the world champion Go player, Lee Sedol. After his defeat he commented on the machine’s winning move: “I’ve never seen a human play this move. So beautiful.” The game of Go was previously thought to be out of reach of AI but Alpha-Go’s extensive use of ML and its ultimate success shows how powerful this technique is.
The financial sector is pursuing many possible ML applications. One area of interest is fraud detection. The theory is that given enough sample data of fraudulent and honest activity, a machine can learn how to differentiate between the two and thence be of practical value in fraud prevention. Such a system might, for example, by itself discover Benford’s Law. Benford was there first though (in 1938) with his discovery that in many real numeric data sets, numbers starting with the digit “1” predominate, followed by “2” and so on down to “9” which is seven times less frequent than “1”.
Application of this law could easily have caught Bernie Madoff, whose returns data on his funds didn’t match the Benford pattern (because, as we now know, Madoff was just making them up). This highlights another issue in this area. A fraudster who knows about Benford’s Law can easily make up fraudulent figures that comply with it so, in the long run the fraud prevention system needs to be one jump ahead or, perhaps, be incomprehensible in its reasoning – a feature that is often thought to be a disadvantage in AI financial systems.
It is still early days for AI in banking but given the level of interest it is likely that the financial world will see more and more AI applications going live in the near to medium future. But Tomer Garzberg, chief executive of Gronade, an enterprise growth lab and future of work company cautions: “AI isn’t a free-thinking sentient machine. There is not enough computer power in the world for that.” Instead, he says, AI can be seen as a marketing term, an umbrella statement for a group of technologies that sit underneath it.
There are many active areas besides fraud detection, for example, anti-money laundering, hyper-customisation of services and identification of untapped business opportunities. Natural language understanding has a part to play too, allowing non-technical employees to access data analytics insights or combined with voice recognition as a facility in call centres. I will leave the last word to Garzberg: “The way we work is changing. AI is not disrupting tomorrow. It is disrupting today.”