Developments in artificial intelligence (AI) are promising but frequently cause concern that developers may lose control over their digital innovations and have to deal with unintended consequences of substituting human decision-making with algorithm-based analysis.
Professor Neal Bangerter from Imperial College London believes much of the media coverage of AI is misleading hype based on sensational headlines that have very little to do with the current state of the field.
Bangerter told delegates at the Cayman Investment Forum on 27 Oct. that he, too, maintains a healthy dose of scepticism about where AI is going.
He argued that the ability of machines to think or perform cognitive functions that we normally associate with the human mind is not yet as advanced as is commonly believed.
Pattern recognition
AI is fundamentally data-driven and consists of machines processing data in quantities and at speeds that simply are not possible for human beings.
While humans are very good at spotting patterns in small amounts of data, that ability diminishes significantly as the amount of data increases. That is where machine learning comes in.
“We’re not just talking about augmenting things that we can already do as humans and doing it better or faster,” Bangerter said. “I think, fundamentally, we’re gonna start seeing a shift where machines are going to be able to do things that we simply can’t do right now and enable applications that aren’t currently possible.”
Rather than robots taking over, he said, the reality will be “much more mundane” with “an era of rapidly increasing productivity” as machines get much better at performing different narrowly defined tasks.
Replicating human brain is far away
In the 1950s, at the dawn of artificial intelligence, researchers believed that machines as intelligent as humans would exist within a generation. When, or if ever, that is going to happen is difficult to predict, Bangerter said, when pressed during a Q&A session.
“If you look at the technologies that we’ve got available right now and the algorithms that we’ve got available, we don’t have a technology that we can point to that could accurately perform all the functions of the human brain. Given that technology doesn’t yet exist, if I have to hazard a guess, I don’t know if we’re ever going to get to that point. But certainly not within the next couple of decades.”
At this stage, he explained, artificial intelligence is influenced by three factors.
There are state-of-the-art algorithms that allow machines to tackle various problems. There is a massive explosion of data and sensors that constantly generate new data. And there is a much greater ability to store data, as well as advances in computing power.
Yet, at the same time, there are many hurdles that need to be overcome.
Clean data
One of the most important issues is the quality of the data.
“Data is what AI feeds on and if we don’t have the data, the algorithms can’t do much,” he said.
AI requires clean data, which rarely exists in the real world, Bangerter noted. “From an investment perspective, I think the data problem is one of the biggest hurdles that we have. Firms that are really going to emerge as leaders in the coming decade are ones that focus on producing very, very clean datasets that can feed AI and machine-learning algorithms.”
One way in which some AI projects ensure clean data is through supervised learning. A computer algorithm is trained by providing input data, such as MRI scans or electrocardiograms, and output data, for example, describing the presence or absence of malignant tumours or heart defects.
Once an algorithm is trained in this way with sufficient clean data, using this so-called data labelling, it can often spot specific health issues with a much greater accuracy than the human expert.
But the ‘generalisability’ of the data can be an issue. Bangerter has been working on a supervised learning algorithm to diagnose hip dysplasia in infants, a condition where the hip joint does not properly form. Hip dysplasia is easily treatable when recognised early.
It is diagnosed using ultrasound images. The algorithm was trained using thousands of ultrasound images from paediatric patients in Oxfordshire in the UK that were already diagnosed by an expensive orthopaedist.
The algorithm turned out to be 92% accurate analysing new ultrasound images from Britain. But when tested against a sample of ultrasound data from India, the accuracy was nowhere near as high.
This may be due to subtle genetic differences in the size and weight of the babies or the result of the differences in imaging data produced by another ultrasound device.
The challenge is to develop an AI solution that is robust to slight changes in data, he said.
Computing power
There is also still an issue with computing power. In the late 1990s, researchers discovered that many of the mathematical calculations required by AI could be performed by running several graphical processing units, graphics chips used in video gaming consoles, in parallel, rather than using computer processing units.
It restarted the development of artificial intelligence after a period of ‘AI winter’ when much of the theory could not be tested in practice due to a lack of hardware.
Despite the advances over the last two decades, there is still not enough processing power for the latest AI algorithms or to simply train a healthcare AI solution on three-dimensional MRI datasets.
Regulatory and ethical issue
Finally, there is a legal and regulatory minefield to consider together with ethical concerns.
Bangerter says many questions need to be answered before AI can be adopted more widely. For instance, who is liable when something goes wrong using AI-automated health diagnoses?
For applications like self-driving cars, it may take decades, not only to equip all cars with similar sensors but also to put the regulatory frameworks in place.
There are real ethical and responsibility issues that must be looked at in the short term and balanced against the tangible benefits AI solutions can deliver for people’s lives, he added.
Related Videos
