By MH van Staden, Data Analytics Leader of the Year 2020, Speaker, Writer, Mentor and Coach
What is this AI everybody is talking about? Well that probably depends on who you ask. For some the term artificial intelligence entails the cutting edge of self-driving cars. When others say AI, they mean automatically granting a loan.
In theory, the term ‘AI’ does include all of the above, as well as having computers performing even more elementary feats of what we once saw as human only capabilities, like adding 1 + 1.
These days, ready availability to data scientists everywhere, through open source packages, is leading to wider explorations of more layered AI. Despite the illusion that speaking robots might create, even the most complex learning algorithms are however not capable of general intelligence, out of the box.
Point your AI laser in the right direction
Machine learning, the process by which AI learns, generally excels at optimizing how to solve a very specific task. It is thus key to spend enough time between data scientists and subject matter experts, refining the exact modelling ask.
If you’re chasing the best possible bottom line value impact from your marketing campaigns, purely having your models solving for volume of sales, might very well not deliver the results you were looking for.
It is a bit like setting performance objectives, or focusing training on achieving a predetermined objective.
At the same time, models do age.
Keep your models on their toes
Even the smartest AI, learnt on customer behaviour from a decade ago, will probably not perform well in catering for what people are up to or looking for today. During turbulent times, like those we’ve been experiencing with Covid, assessing a person the same way we did a month ago, might not even be relevant now.
Automated decision making systems need to be monitored closely, on a regular and ongoing basis. Tracking the expected business outcome, like take up or default rates is a good start. Ideally you also want to keep a close eye on the inputs, such as income levels, driving the decisions.
Even this is not enough.
Ask your AI to explain itself
AI has been around for a while. In the early days, developing computer abilities to draw intelligent conclusions went hand in hand with a clear audit trail of logical reasoning. As machine learning approaches have evolved, towards more accurately solving more complex challenges, it has become more difficult to explain exactly what is happening behind the scenes.
In striving for greater predictive accuracy, explainability has taken a back seat, but some methods have been developed, towards interpreting ‘black-box’ implementations as a whole, as well as individual outcomes.
Superior modelling of real world phenomena does thus not need to come at the expense of transparency. As a non technical stakeholder, you’ve got every right to demand a clear understanding of how and why certain automated conclusions have come about.
Gut feel still has a critical role to play in sense checking.
AI can be wrong or right
Sometimes AI-driven decisions are spot on. At other times, not.
There exists a wealth of potential, in utilising advanced algorithms, to map extremely involved processes, in the service of taking better actions off the bat. Whilst we’re at it, deep learning, or layered machine learning, can help us discover things about our customers and organizations we never knew before.
Realising these benefits, necessitates asking difficult questions:
- What exactly is your AI solving for,
- How well is your AI doing that &
- Why is your AI coming to specific conclusions
Until now, it has largely been those technically involved, just getting themselves comfortable enough with what they have done. For improved AI acceptance, theoretical tools developed will need to be applied, to explain more clearly, to a wider audience, what exactly is going on under the hood.
This requires a combination of retaining sufficient data science knowledge and incubating constructive cross functional dialogue.
Successful AI requires interpersonal skills.