G'day, I'm Lee Hopkins

Today I'd like to talk about three things: ways to think about machine learning, AI and Customer Service, and ethics in AI.

Benedict Evans has written about machine learning and ways to conceptualise it. Five years into the serious use and application of machine learning and we are still no closer as a society to be able to simply explain what machine learning is.

As Ben says, "This isn't helped by the term 'artificial intelligence', which tends to end any conversation as soon as it's begun. As soon as we say 'AI', it's as though the black monolith from the beginning of 2001 has appeared, and we all become apes screaming at it and shaking our fists." 

Ben suggests we think in terms of relational databases. Back in the 1960s and mid 1970s, before relational databases appeared, if you wanted to query your database you had to write a specific query for it. You couldn't just ask for a list of all customers who had bought from us in the last six months and who lived in Adelaide and expect the database to whip up an answer. Coding had to happen. 

"Databases were record-keeping systems; relational databases turned them into business intelligence systems — Ben Evans"

Then along came relational databases and with them the opportunity to ask questions on the fly. But databases, even relational ones, were only as good as their data. The same with today's machine learning algorithms. 

And today's algorithms don't cross-pollinate. Just because IBM has Watson, it doesn't mean that every other AI/ML system will get smarter when Watson learns something. Just because one company enjoys success using machine learning to cut costs or increase sales, it doesn't follow that other companies running ML algorithms will be similarly successful. Not all attempts to harness AI will be successful, or as successful as each other.

As more and more marketers, and the businesses that support them, venture down the AI/ML path, there will be wild claims about this particular algorithm's efficacy, or that particular algorithm's cost-effectiveness. As in all gold rushes of the past, buyer beware.

The uses of AI and ML are still being discovered. Sure, there's some low-hanging fruit, but the 'so what next?' questions will continue to entertain us for the next 10-15 years.

Ben has a weekly newsletter that is well worth subscribing to if AI and ML are an interest of yours.

https://www.ben-evans.com/benedictevans/2018/06/22/ways-to-think-about-machine-learning-8nefy

AI and Customer Service

Picnic is a fast-growing online supermarket. Based in the Netherlands, it uses Natural Language Processing to help remove some of the drudge work involved in Customer Service (or 'Customer Success' as Picnic calls the team), and speed up resolution and Customer Success staff service times. 

Dictionary.com calls NLP, "the application of computational techniques to the analysis and synthesis of natural language and speech," and Wikipedia says, "Natural language processing (NLP) is an area of computer science and artificial intelligence concerned with the interactions between computers and human (natural) languages, in particular how to program computers to process and analyze large amounts of natural language data."

As Bas Vlaming points out, "Natural Language Processing is a very broad area, and concerns the interaction between, on the one hand, language as used by actual people, and on the other hand, computers, automated systems, algorithms and the like. This is, unsurprisingly, a very non-trivial task. Natural languages may be ambiguous, their structures, syntaxes and vocabularies have changed throughout history in ways that may not always make sense from a purely logical perspective."

So Picnic is on the money in their decision to use NLP to analyse the messages they receive from customers via customer satisfaction phone apps. The customer service app can send images and text, but at this stage, only the text is analysed for further processing. 

Picnic's approach to product feedback consists of four steps:  1. Identify what messages concern product issues 2. Identify the type of product issues 3. Identify what product and the order the customer is referring to   4. Suggest appropriate resolution, and allow this to be executed directly

This last step is carried out by human Customer Success people, but the three prior steps can now be carried out by computer, saving the busy Customer Success person minutes per issue.

You can read Bas' great post over on Medium, or on Picnic's own blog. 

https://blog.picnic.nl/machine-learning-and-customer-success-a-beneficial-partnership-ed2aab66dc71

Ethics

Martin Eve points to the incongruity and mismatch of ethical decision-making when training algorithms to pattern-recognise, or even just recognise a photograph. 

Claiming that US billionaires are routinely paying pittances to the lowly staff who feed data into the computers, Eve reminds us that sorting through child abuse imagery for hours at a time, for around $0.02 cents an hour, is scarring for the worker and potentially trauma-inducing. 

It is the lowly-paid researcher that has to do the donkey work in most research labs. In Google and Facebook that work is outsourced at two cents per hour. 

Let us not forget, when we eventually create machine learning algorithms in our own companies, that the data entry operators need a living wage too, to compensate for the potentially traumatic work they are undertaking for our profit.

https://www.martineve.com/2018/07/02/the-real-ethics-of-AI-are-about-the-labour-underpinning-it/