The basic premise of artificial intelligence, to use enormous amounts of data to find out new things, is easy to grasp. If any one of us had the time and stamina to study a million photos or stories about a thing, I’m sure we’d come up with insights about it too.
Business products emerging from current applications of artificial intelligence are also logical and simple to get your head around. Smart thermostats sell because they are convenient and deliver energy savings. Marketing approaches that analyze shopping patterns to suggest items people are likely to buy are winners in retail for their potential to increase sales.
How does AI get from data analysis to creating desirable products? In diagram version, this seems to me:
A few hypothetical1examples:
1. Using AI to improve diagnosis of medical images. Input: One hundred thousand pathology slides of renal cancer and one hundred thousand slides of normal kidney tissue. Outcome: Improved differentiation between normal and malignant kidney biopsies. Doctors win because the accuracy of diagnosis increases, saving healthcare costs by prescribing the right treatment for patients. Patients win because they are either can carry on their lives disease-free or have greater certainty in the treatment they need.
Mysterious GUIey2inside: What is the AI looking at to distinguish between normal and cancerous cells in pathology slides?
2. Using AI to improve traffic flow. Input: Every car in the city communicates its starting point, destination, and real time location to a central database. The goal is to send a uniform volume of traffic via every available route so that none are over-used or under-used. The outcome is a clear win – optimum travel efficiency for everyone, saving time, auto costs and impact to the environment by decreasing energy consumption.
Mysterious inside: What is AI doing to manage all the permutations and combinations to direct even traffic flow?
The two examples are different. In the first one, the criteria AI uses to distinguish between normal and malignant cells are the mystery. Pathologists could list the traits they use to make a decision when looking down a microscope, but is AI using the same ones? In the second, it’s the speed and capacity to deal with volumes of users that’s amazing. It’s not difficult to suggest the best route for your mother to take home, based on knowledge of traffic patterns at the time of day in your home town, but who could do that for 3 million occupants of a city simultaneously?
I’ve read that we are unlikely to be able to extract the GUIey middle3from AI supported processes, due to the iterative nature of the learning. When a person really understands what they are doing, they can explain it. If a chef tells you their sumptuous meal resulted from ‘a little of this, a little of that’, they likely know exactly what went into the dish, but aren’t telling to protect their trade secrets. If my mechanic tells me they are basing the diagnosis of what’s wrong with my car on some data from other cars but doesn’t know which models or what kind of data, I’m looking for another mechanic.
Is not knowing how AI works any different than not knowing the detailed working of automobiles, or any other complex object or process in modern life – elevators, mortgage documents, dental implants? The fundamentals of the car I get – the energy of exploding fossil fuel is converted into angular momentum that torques the axels and moves me, in my steel and plastic carriage, to where I want to go. The business model is also easy – the speed and convenience of reaching destinations in relative comfort with the added efficiency of carting a group of people, sheets of drywall, or my dogs with me. There is someone who can explain ABS brakes, how the muffler is connected to the engine, and all the other components that make a car function. With AI, either by design or trade secret, the explanation is hidden.
We need to know the mysterious processes that AI systems use to derive new knowledge from the volumes of data consumed. Forget proprietary algorithms. This is brave new territory we are entering and transparency is important so we can be sure we are operating safely and ethically.4
History is full of examples of embracing new things without a full understanding of the implications5. From that, a machine would learn that we need to know how things work before we can use them safely.
1Both of my examples are likely to be real enterprises but staying hypothetical is better for this discussion.
2This is a pun on GUI type computer interfaces, which use icons, rather than typed commands, to tell computers what to do. GUIs make programming simpler. I’m suggesting by making things simpler with AI, we are making them less transparent, dissectable or amendable to understanding how the parts work together to create the whole. Less concrete. More gooey. Gooey-er. Soft and flowing, changing shape easily.
3I do know that the process AI uses is a very large series of logic functions, of the sort: if X does Y, then A is the outcome. If X, K and J, do B, then L is likely to happen. If X does Y but K does something else, and it’s Tuesday, then Blue is the right answer. Etc. Oh, and the AI may start with a bunch of logic statements but change them on the fly as more data comes in or if in testing a hypothesis, it doesn’t deliver satisfactory answers.
4For many examples, read ‘Weapons of Math Destruction’ by Cathy O’Neil
5A few examples that spring to mind – nuclear weapons, cigarettes, social media, plastic, many types of home insulation, lead paint, breeding of dogs, trans-fats, mortgage backed securities.