Artificial intelligence is a term that is thrown around a lot. Knowing what it really means allows marketers to better assess its opportunities, or threats. We asked RayPoynter, founder of The Future Place, and of NewMR.org, to demystify some of the core elements of ‘AI’.
The 1968 movie “2001: A Space Odyssey” questioned the role of artificial intelligence and highlighted potential fears. Artificial intelligence is all around us today, but most people seem to be overestimating or under-estimating the changes it is creating.
The term artificial intelligence, or the friendlier AI, has become an umbrella term for a wide range of approaches and tools. Attempts to define AI too tightly are a pointless distraction. Most vendors of software systems/platforms (from advertising delivery platforms through to the latest washing machines) claim to be utilising AI. Rather than engage theses vendors in wasted time discussing whether something is or is not AI, it makes more sense to investigate what principles their system uses and the capabilities and limitations of their approach.
There are broadly four ways of approaching artificial intelligence, these are:
- Supervised Machine Learning
- Expert Systems
- Unsupervised Machine Learning
- Deep Learning
Supervised Machine Learning
Supervised Machine Learning is currently the most widely applied AI approach, including such uses as Google Translate, the coding of images, and predicting sales, preferences, or churn. When somebody refers to just ‘Machine Learning’ they are usually talking about Supervised Machine Learning. The learning in the name refers to the way that the ‘intelligence’ is created. The program is given a set of inputs, for example, a set of social media comments, and a set of outputs (for example the result of a human having coded them as positive, negative, or neutral). These two elements are called the training set. If the computer can create a set of patterns that enable it to correctly predict the outcomes from the inputs, it can then be used to predict the outputs for other inputs, i.e. it creates a predictive model.
Key limitations of Supervised Machine Learning are:
- It needs a training set, typically this means thousands of inputs, described in clearly defined variables, with clearly defined outputs. We might imagine Amazon doing that in terms of what purchases predict a purchase of product X, but it would be hard to imagine (with current technologies) defining what variables fully described a TV commercial and then matching thousands of them to unambiguous results.
- It only works if the variables in the input set actually predict the outcome. If there are missing variable, or if the outcome has too large a random element, then machine learning won’t provide a solution.
- The solution only works when used in the context of the training set. For example, if we trained a system to look at the language of marketing for travel, we might find that ‘risk’ was a negative term, but applying that solution to financial products it might miscode many entries – because ‘risk’ plays a different role.
Expert Systems are almost the mirror opposite of Supervised Machine Learning. In an Expert System, the computer is given a set of rules and then applies them. For example, it is possible to write a short set of rules that describe the best way to play a game of noughts and crosses. The main problem with Expert Systems is that there have to be rules and the rules have to be known. The advantages include not needing to have a large training set and knowing how the system works. For example, AI systems that write articles based on company financial reports are typically based on Expert Systems.
Unsupervised Machine Learning
Unsupervised Machine Learning refers to situations where there is not a training set. The machine is given a set of input data and is asked to discover patterns. A cluster analysis program is a simple example of this approach. For example, a program might be given a book as an input and asked to identify themes and patterns. In most cases, this sort of program works in conjunction with an analyst who may seed the system with some ideas, and who will take the initial outputs and impose additional rules to help produce a useful result. Supervised Machine Learning and Expert Systems can produce a product that can be used as a tool – Unsupervised Machine Learning tends to produce a specific result to a specific problem.
Deep Learning is a system where the outputs from the model are fed in as new inputs. Deep Learning had a burst of fame when Google’s AlphaGo beat the world’s best Go player. The first step was to teach the computer to play, in effect using an Expert System approach, then the computer is fed a large collection of games between strong players, creating a training set (i.e. Machine Learning). The next step was to set the computer to play itself, using different strategies from its learning. The computer used the results of these games to learn how to play better. Deep learning has the potential to improve processes where a large number of specific actions can be tried and the results are known – for example, book or film recommendations and promotions. Deep Learning can’t be allowed to ‘play games’ in order to learn.
What does the future hold? We will see more and more AI. In the short term, we will see a mixture of Expert Systems (writing reports, designing campaigns, creating buildings and engines) and Supervised Machine Learning (wherever a training set can be identified, for example, translation, coding, and data-led marketing). In the longer term Deep Learning and beyond they General Artificial Intelligence will change things even more. General Artificial Intelligence? Ah, that is another topic worth writing about.
If you’d like to know more about AI (and automation in general), NewMR.org are holding two webinar sessions on 20th September, with eight presentations – Sign up here.
“Just what the industry needs, great collaboration between clients and agencies on the topics that drive business growth.”
Bridget Angear, Joint Chief Strategy Officer at AMV BBDO
“It’s great to see the IPA in the UK bring the whole industry and particularly the trade bodies together to focus on effectiveness. This new Marketing Effectiveness initiative will enable people across the industry to work together to build on best practice.”
David Wheldon, Chief Marketing Officer, RBS
“Effectiveness is a team sport, so it was great to see the industry in the widest sense, come together. In an increasingly diverse and fragmented world, only by using all parts of the brain will we solve effectiveness challenges and design our campaigns to deliver short and long term value. That’s why what happens next is important – if the IPA can help facilitate progress on this with a long-term initiative around Marketing Effectiveness, we’ll definitely crack it.”
Bart Michels, Global CEO Kantar Added Value and Country Leader Kantar UK
“The time spent at #EffWeek was extraordinarily effective. It was great to hear the diverse views from all areas of the industry. All tied together with the common themes of accountability and effectiveness.”
Andrew Canter, Global CEO, BCMA
“It has been a privilege to be part of the inaugural Effectiveness Week. The agenda is one which we at O2 UK feel passionately about. To see and hear perspectives across the industry demonstrates how the breadth of marketing effectiveness is increasingly being valued within businesses. Data, insight, social, customer experience, test and learn, ROI, these are all fundamentals and were covered expansively at the event”.
Sandra Fazackerley, Marketing & Consumer, Telefónica UK Limited
“The full week of effectiveness events brought into clear focus the need for marketers to use data and insight to achieve the key business objectives of growth and profits. Marketers today are in a better position to quantify their knowledge of customers and measure the ability of investments in marketing to increase brand and shareholder value.”
Chris Combemale, Group CEO, DMA