Human Intelligence AND Artificial Intelligence

Let us assume you have developed the best industrial AI ready and want to deploy to the shop floor, now what to expect?

The AI you developed predicts nicely the process, identifies causal impact of different regressors and even provides optimal control. You install it on a machine on the shop floor. You get back a while later to check. So far so good?

Not really: expect the AI to be offline because the machine operators switched the system to manual control.

More often than one would think, Human Intelligence (HI) and Artificial Intelligence (AI) do not go hand-in-hand: it is more HI vs. AI.?

I described a realistic, although oversimplified scenario, I agree, but even in this form the scenario exposes the fact that AI must gain acceptance on its own merit. And it is an uphill struggle. No red carpets in the name of AI.

The lukewarm acceptance, sometimes outright refusal of an external controller is well known in literature: it is called the out-of-the-loop syndrome, meaning that if you are not in the decision loop, you hesitate to accept decisions. More so when you are in front of a black-box which is typically the case for conventional AI, think Convolution Neural Networks (CNNs).

On top of that, people in operations - including machine operators, supervisors, managers - are measured by their KPIs, typically some form of output-per-time-unit, and some part of their salary depends on it. So they are rightly careful about a new technology that may improve the company’s productivity, but also make their KPIs worse off.

In their shoes, I would also ask for hard proof on the part of AI before accepting AI to run the process for me.

How can you provide a performance proof for AI on the shop floor? The question is more tricky than it looks at first glance: you may run simulations or you may even run actual AI tests on a pilot machine and measure large benefits. But would those results apply to your other similar machines in the shop floor? You cannot blow your chance because operations are always so tight with allowing tests that may endanger output...

The short answer is: there is no guarantee you will get similar results.

The longer answer is that scientists have a name for this question: it is called the transportability or external validity question. They have worked out the science to tell us when external validity holds. In other cases, you are on your own and machine operators are rightly skeptical.

Besides AI acceptance, the external validity problem may sound like a dusty academic research theme, but is not. External validity has big implications: if the learnings from one machine can be somehow transferred to another machine, this also means that they can be cumulated meaning that your AI will get better the more machines it runs on.

One approach with proven results in external validity is Causal Inference, which has one more key feature: the causal AI models are depicted graphically, meaning they can be shared with subject matter experts early on to develop on top of their expertise and sharpen their assumptions when data requires.?

At Aleph we embraced Causal Inference because it gets people involved early on, which is a pre-condition to have acceptance of AI. And a first step in the journey towards HI and AI working seamlessly together, which I posit is the real goal of AI: giving people super powers so that people can focus on what they do best.