Machine Learning & Expert Systems: they’re not necessarily the same thing

In 1973 I considered ‘Artificial Intelligence’ as a postgraduate research degree: as it worked out, I stayed with Organic Chemistry.  In the intervening decades it seemed to me that not much happened with artificial intelligence, except hype and hijacking “AI” to perhaps mean Augmented Intelligence, Advanced Innovation or similar. Genuine AI advances were happening, but these quietly passed me by, as had Blockchain, apart from crypto-currency.  In April 2019 I went to the combined Internet of Things (IoT) / Artificial Intelligence / Blockchain / Cyber Security expo at Olympia, London.  Genuine working implementations were on show, not hype nor prototypes.  Industrial IoT (IIoT) applications were on show, not tabloid-dumbed-down ‘your fridge will order your milk and yoghurt’ uses.

I had been sceptical about current claims about artificial intelligence as being no more than programmed ‘expert systems’.  Adrian Hopgood (Professor of Intelligent Systems, University of Portsmouth, UK) is gently educating me to be careful about sloppy thinking on my part on the nuances here.  Here’s an example that explains what I think is the difference between artificial intelligence and a programmed expert system:

A blood test interpretation program might pick up high levels of the blood ketone β‑hydroxybutyrate (BHB) and alert as ketoacidosis – a dangerous condition for diabetics.  There is however an alternative, depending on the blood glucose levels: high glucose too would indeed support the ketoacidosis diagnosis.  A low blood glucose level with high ketone is however indicative of ketosis – this is considered a good metabolic status for those avoiding carbohydrates in their diet.  An expert system not programmed with this diagnosis could mislead a physician.  An AI based system that has been trained on ketoacidosis data might flag ketosis if it had also been trained with this combination, or an anomaly if it had not been exposed to this set of results.  It’s this uncertainty, entertaining an unknown alternative, that differentiates the artificial intelligence approach from a programmed expert system.  A medical AI expert system not ‘aware’ of ketosis is a safer utility than a programmed expert system that has not been programmed to recognise ketosis: a confidently wrong diagnosis could be positively dangerous if the doctor has excessive confidence in the programmed expert system.

Artificial Intelligence comes in several different forms and categories.  The above example is numerical data with supervised training: the AI has been told “this is what ketoacidosis looks like”.  Voice recognition is another form, which needs to overcome the foibles of accents and homophones.

An industrial application of AI using picture recognition currently in development at the University of Portsmouth led by Jiye Chen, is to identify broken solar panels on a solar farm.  The existing monitoring system can detect the number of solar panels not working, but not which individual panels.  With hindsight, perhaps this could have been included in the Service Design phase.  The proposal is that a drone flies over the solar farm: the feed from its camera is sent to an AI system that determines which solar panel it is looking at.  This is not an easy question for AI: “Is this a solar panel?”.  Determining the panels’ functional state “is it working” is relatively easy where the camera feed extends into the infra-red, as is the panel’s location.  Since solar farms can cover large areas, with the largest in England at Chapel Lane Solar Farm, Dorset covering 310 acres / 125 hectares, a person on foot might take several days to complete a survey.  While alternative technologies might also work, this AI approach has been assessed as potentially the most efficient.

AI also comes into its own when there are vast amounts of data available and speed of response is critical.  A suite of network protection products from Darktrace utilises unsupervised machine learning.  Over four weeks in what is referred to as ‘unsupervised learning mode’ the appliance determines what is normal in an enterprise’s computer network.  In autonomous mode it can lock down a potential threat in less than 2 seconds.  That may be a genuine threat like ransomware attempting to deploy, or a spurious correlation like the high-confidence 99.8% match for 1999 to 2009 data for “US spending on science, space & technology” with “US suicides by hanging, strangulation and suffocation”.

My car sat-nav is definitely not AI nor even AI-like: many months after a roundabout junction near where I live has been closed, with a new by-pass road opened, and repeatedly driving on the new road, the sat-nav insists I’m driving over a farmers’ fields.


View comments on this article on LinkedIn: