Can We Rely on Machine Learning for Critical Applications?

I am no stranger to Machine Learning, since I work for an AI company, which last year brought several machine-learning (ML) applications to various industries globally. This doesn’t surprise me as INFORM has been named one of the 20 most promising AI companies in the world and, most recently, one of the most innovative companies in Germany. What did surprise me, though, is a more personal experience with AI, specifically machine learning, when I caught myself trying to figure out why there was dog poo in my home office.

Quick backstory… I’m a big fan of robotics and smart-home solutions. I’ve had a robotic vacuum and lawnmower for some time now, and over the summer, we introduced a second, more advanced robotic vacuum to our home arsenal. This specific vacuum was very high on my “want” list because it is quieter than our original vacuum, but also, it came equipped with machine-learning capability.

Out of the box, the vacuum delivered on its promise of noise reduction. Another unique selling proposition of the vacuum was its double, front-mounted camera system, which allows it to detect obstacles like toys, cables, shoes, etc. Unlike its predecessor, this robot didn’t just mindlessly use a bumper sensor to detect an obstacle in its cleaning path. The vision system allowed it to avoid the collision altogether. Before long, the system started logging interesting obstructions to its cleaning pattern, like switchboards, dustpans, and to my surprise, dog poo.

So, when my new robot informed me that there was dog poo in my home office, I was rather perplexed. You see, the problem was less the content of the alert and more the fact that I don’t even have a dog. That would mean I would not only have to clean it up, but I would have to figure out how a dog got into my office in the first place. To my relief, when I entered my office with gloved hands, paper towels, and cleaning supplies, I was relieved to find nothing I should be afraid of.

I decided to do some digging to better understand what might have happened here. After all, I’m not a big fan of being told there is dog poo in my house. What I found was that the machine-vision system in my new robotic vacuum uses machine learning to continuously learn what objects are. Using a supervised machine-learning technique called “classification,” the AI system has been, or rather is slowly being, trained to determine different types of obstacles it may encounter. As the knowledge base and experience of the algorithm grows, so too will the accuracy of the predictions on obstacle classification.

In the case of the mystery outcome of cleaning my office, I was able to download a picture of the obstacle – the result? The vacuum had “seen” the black base of my office fan, and its machine-learning algorithm determined with an 83% probability that it was dog poo. To avoid spreading it while cleaning, it flagged it in the app and avoided the obstruction.  Given the vacuum is taking a unique decision each time and constantly learning, I’m confident that the issue will resolve itself with enough feedback that my fan is a just a normal fan and is not sitting on a pile of dog poo. But it did get me thinking more broadly about what the state of machine learning in the logistics sector is and whether we’re any further along than my new robotic vacuum cleaner.

Even today, there is a wide range of existing AI and ML solutions surrounding logistics within the market. Global supply-chain-planning solutions analyze vast sets of data, and applying smart algorithms allows organizations to balance supply and demand while concurrently optimizing the delivery processes. Diving deeper into the logistics network, down to a single location AI and ML applications exist that predict demand and improve demand forecasting. They also monitor and predict traffic and other factors that may affect the quality of your transport or shipping times at your site. And finally, digging even deeper into the micro-cosmos of sorting-operations, robots driven and enriched by AI and ML are likely to soon be helping human employees to unload or load trucks and containers. When we get to this level, the comparison to my robot vacuum becomes overtly clear.

So here I sit pondering the question, “What if something like this happens in real-life operations of a logistics company?” Given our increasing reliance on AI and ML to support the operations of logistics operators around the world, I think it is a pretty good question to be asking, and when one does, they come to logical follow-on questions, such as: What happens if the technology makes an incorrect prediction? Who is responsible for the potential financial consequences or service-quality impacts? Are the AI and ML solutions available today reliable enough to support operators? In the future, how do we provide acceptance signoffs on ML solutions if, by their very nature, they are meant to continue to evolve and learn? How do you know when you’ve got it right?

I know these are a lot of questions without answers. That said, sometimes it helps to pause and ask questions you don’t intuitively know the answers to. In the coming months, as I ponder these questions, I’m sure more will arise, and perhaps I’ll find some answers that might fit. As I do, I will come back to this topic and provide updates. Until then, may there be no surprise dogs visiting your home office.