Achieving balance in Machine Learning development
This is a guest post for the Computer Weekly Developer Network written by Ian Greenhalgh in his role as UK & US managing partner at Delaware.
Spelled in lower case, delaware is a software systems consulting organisation with specialist skills in SAP S/4HANA and Microsoft technologies — the company works across multiple industry verticals including automotive, chemicals, food, retail and the wider professional services sectors.
Greenhalgh contends that employees today in modern forward-thinking organisations are increasingly frustrated that the ‘intelligence’ available to them in their personal life isn’t available to them in the workplace
But, he says, his company is seeing technology being applied in a number of ways to help counter this shortcoming.
Greenhalgh writes as follows
Intelligence is happening. Robotic Process Automation (RPA) technologies and software-based bot’s are being implemented to automate mundane repetitive tasks, many of which were traditionally performed by service centres.
RPA is also being used to automate the handling of process exceptions, the reason being that it’s often cheaper to automate an unwanted process variant than to try to eliminate it.
Another burgeoning use case is the identification of emerging issues within large operational datasets, Machine Learning (ML) is triggering alerts to managers highlighting potential problems within a production process or a supply chain — issues that the user might not have spotted unless they looked for them specifically.
This requires AI specialists to work with client process experts to identify a set of problems and then create the necessary training, development and test datasets which can be used to train and prove the ML algorithm. The challenge is to identify concrete data sets that allow the algorithm to recognise the exception from the general process noise. In cases where relatively large numbers of problems exist, producing distinct and differentiable datasets can be difficult.
Steady, but solid (learning)
The approach is very much exception by exception, business issue by business issue, but the steady result is that the algorithms monitor increasingly more process exceptions.
Evidently, early problem identification is a huge opportunity, but a balance needs to be struck. If the machine hasn’t been trained to deal with certain problem types or patterns, an overly reliant user might not be alerted to the problem until it is too late.
The answer to whether machines could ever completely replace humans is, unsurprisingly, sometimes.
But, we must stay vigilant of the two biggest issues; getting humans to trust the results provided by the machine and managing (and understanding) the consequences of failure – although 80% of Netflix content is selected by algorithmic recommendations, the consequence of failure is very limited, whereas if an algorithm fails within an airport scanner or a CAT scanner, the consequences can be fatal.
Let’s think about the real world application of ML and how it will affect us and keep automating.