Robot safety part 2: man smart, machine smarter
Machine learning seems to be a new buzzword. What is it and how is it substantially changing the way we work? John Kersey continues his series on artificial intelligence in the world of health and safety.
Perhaps the first application of machine learning was the Checkers Player program devised by Arthur Samuel in 1952, with the 1955 version incorporating machine learning. This first ran on an early IBM 700 series computer, by 1962 the checkers master Robert Nealey took on an IBM 7094 which subsequently defeated him. How was it able to do this?
There are different ideas on machine learning and broadly two approaches. The cryptographer and famed Enigma codebreaker, Alan Turing in 1950 thought instead of devising a system where everything was laboriously programmed it would be better to devise a “Child Machine” that would when “subjected to an appropriate course of education one would obtain the adult brain”. The Checkers Player program would run through many thousand variants of the game and find the optimum solution.
These ideas are the essence of machine learning. Besides these variants in approach there are differences in methods used. Machine learning can said to be either “supervised learning” where human first “trains” the machine in which response is the desired one against set data or alternatively “unsupervised” where there is only input data.
In a translation example, you could use a text in the original language, and the same text in the target language, to carry out supervised learning. In most safety applications we will be dealing with unknowns or a great variety of variables – so how do we deal with that?
Fortunately, statistics gives us several techniques which the machine can employ to make sense of the data offered. One example is clustering, where similar information might be clustered in a group. In an occupational road risk application, we might find a cluster of road traffic accidents clustered around alpha management in high status vehicles, in clear conditions on motorways, around service vans, long working hours, in extreme conditions on minor roads. Clearly these will require different strategies to control, but they can be tailored to the insight.
The real power of machine learning over human learning is most normal humans will see links in one or two factors whereas the machine can compare factors over a wide range of datasets in various forms. With cognitive systems like IBM Watson it is able to compare and cluster information on individuals such as personality through social media feeds or text mining, sensors such as telematics in vehicles, and conventional datasets such as claim histories or accident records.
Sometimes these insights will confirm what we expect or conventional wisdom. Sometimes the insights will be much more covert, and see links we do not. Sometimes known as the “beer and diapers” syndrome after an urban legend that guys buy diapers (nappies) for their heavily pregnant wives and get beers as a reward!
Canny retailers thus put the beer and nappies together for extra sales. Using machine learning, a study could predict potential suicides up to 2 years in the future with an accuracy of 80% . This compares with the current 50% level by clinicians (ScienceDaily 1/3/2017). Some examples of safety case studies informed by machine learning from Space Time Insight (www.spacetimeinsight.com) include insights into:
- on-the-job injuries for a major US railroad operator referenced to time and location
- avoiding incidents caused to workers carrying out work on a wind farm by predicting weather patterns
- an examination of the link between increased traffic for a global logistics organisation; doubling the activity does not lead to doubling the accidents.
Deep Learning is another related learning concept. It is based on neural network learning, which models the way the human brain thinks. The computer system “thinks” in layers; the first layer identifies the approximate object for example and subsequent layers “think” more deeply and in more detail.
In this way a Google system demonstrated by Andrew Ng and Jeff Dean could discriminate between cats/human faces/yellow flowers/other objects in 10 million YouTube videos. These could be sorted into any one of 22,000 categories without human intervention. At first the system had a 16% success rate in correctly categorising objects however this jumped to 50% accuracy when the categories were reduced to 5,000.
In many cases the images sorted would have been challenging for humans to categorise. A lesson here is computer systems evolve and can be capable of radical improvement. As Jeff Dean summarises: “Deep Learning is a really powerful metaphor for learning about the world”.
It’s a truism in the computer industry that 90% of business data remains unanalysed and so the rich insights it may hold remain undiscovered. Could it be that by tapping into this rich motherlode of data the health and safety professional could gain a deeper understanding into why things go wrong, or why they go right?
Now machines enable us to gain this insight over a variety of media limited only by our access to valuable data.
Disclaimer: The views expressed in this article are those of the author and do not necessarily represent those of any commercial, academic or professional institution the author is associated with.
Coronavirus advice for employers
This hub page complies all the latest government coronavirus updates. It includes what you can and cannot do in England, Scotland, Wales and Northern Ireland, workplace advice from multiple sources, including information on welcoming staff back to the workplace and the latest vaccination information.
It also contains a host of useful external links and resources to find further information.