The hybrid approach to true knowledge

Connections that computers make can only get better when they use a hybrid of both the top down approach of traditional Artificial Intelligence and the ocean-churning approach of machine learning.

Over the last few weeks in this column, I have written some pieces on how the “cognitive” functions in Artificial Intelligence (AI) would be dead on arrival without correct, contextually labelled data. The first dealt with teaching AI machines common sense, which is a non-trivial undertaking. The second was a complex piece which dealt with how these data needed to be analysed with the use of deep science so that they can be used in the world of medical AI. Last week’s piece dwelled on how this labelling of data, from the mundane to the arduous, will create a large employment opportunity for humans in the Intelligent Age.

For many years, the pioneers of AI stuck to an approach where they tried to capture the best of human “expert” brains to create heuristics—or rules of thumb—that could act as reference points for traditionally built computer systems to use. Many of these heuristics were codified using frameworks which rely heavily on the cognitive psychology behind how humans think and learn. I have spoken of CREDO, an attempt to codify such knowledge using human cognitive psychology which was created by John Fox, a professor of Engineering Science at the University of Oxford. Also, earlier in my career, I worked for a firm that was focused on this sort of AI, and the importance of handcrafting and curating such knowledge stores of heuristics was paramount.

Traditional software development has become notoriously expensive and labour intensive. Managers and marketers who are responsible for procuring and selling software products and services are attracted by “machine learning”, the idea that future computer systems will “program” themselves: learning how to carry out tasks and progressively improving their performance through “experience” without the need for human intervention.

My opinion is that this level of faith in a soulless, repetitive machine is puerile; nonetheless, I will concede that there have been instances where churning through reams of data has allowed computers to make some connections on their own. But these connections are yet unequal to the task—and can only get better when they use a hybrid of both the top down approach of traditional AI and the ocean-churning approach of machine learning. This hybrid approach can be likened to the turbo charged motor car—where the fuel is the set of heuristics or expert knowledge which when fed into computer programmes give off a data exhaust which can then be re-harnessed and added to the fuel to turbocharge the computer’s motor. Starting without the heuristics is like starting a motor car without fuel.

The froth ensuing from marketing departments at some of these AI firms was unfortunately allowed to take over the messaging about machine learning to the public. A few years on, it has become painfully apparent that many of these marketers’ tall promises were damp squibs. IBM, for instance, has faced harsh criticism from some sections of the fourth estate in the last few months. Many articles lambast the claims that the firm’s marketing department has made about the medical capabilities of its Watson product. This disenchantment is also seen among the medical community. My brother-in-law, a professor of medicine at one of America’s well-known universities, and another old friend and classmate, a practising surgical oncologist in Florida, would just a few years ago opine that Watson would save the world by providing pinpoint diagnostic accuracy, and would rid the world of cancer. They were swayed enough by the marketing to tell the doubting Thomas in me to pipe down when I pointed out what I felt were the limitations of a machine learning approach. Both these doctors are now a lot more circumspect when they discuss the impact of AI in medicine. It appears that these physicians are as gullible as you and I are, but at least they are willing to revise their hypotheses when presented with new information.

STATnews, a publication focused on stories about health, medicine, and scientific discovery, is widely read by doctors and has recently published a scathing article on IBM’s Watson, which no doubt contributed to my classmate and my brother-in-law’s change of heart. STAT takes its name from medical parlance. The command “stat” is barked out by doctors to their underlings when they want something urgent done immediately.

The STATnews article doesn’t mince words. The publication alleges that the interviews it conducted suggest that IBM, in its rush to bolster flagging revenue, unleashed a product without fully assessing the difficulties of deploying it in hospitals globally. It further alleges that IBM has not published any scientific papers demonstrating how the technology affects physicians and patients. As a result, says the publication, Watson’s flaws are getting exposed on the front lines of care by doctors who say that the product, while promising in some respects, remains undeveloped. The piece quotes a South Korean cancer specialist who has used the product, who says “Watson for Oncology is in their toddler stage, and we have to wait and actively engage, hopefully to help them grow healthy.” If machine learning systems must in fact start as toddlers, then we need to turn to traditional educators like Fox to train them in the hybrid framework I have discussed earlier in this column. 

OpenClinical, where Fox is chief scientific officer, is a non-profit foundation which seeks to handcraft machine executable AI knowledge from traditional clinical guidelines and research sources. OpenClinical “publets” are accessible in a repository that can be used by medical professionals to create practical services at the point of care. 

OpenClinical is now taking the hybrid approach, introducing machine learning to exploit the “data exhaust” produced by the point of care services to progressively improve clinical decision-making. 

IBM has been teaming with firms and organizations such as these. Its alleged marketing hyperactivity aside, IBM’s willingness to use these sorts of methods to make Watson more robust should be lauded and efforts like OpenClinical and its commercial partner, Deontics, need to be encouraged if we are to see computers playing a meaningful role in healthcare.

Siddharth Pai is a technology consultant who has led over $20 billion in complex, first-of-a-kind outsourcing transactions. He now works as an advisor to Boards, CEOs, and investors to help them strengthen and execute their global technology strategies in an increasingly uncertain and volatile world.

*This article first appeared in print in the Mint and online at www.livemint.com

For this and more, see: