Machine Learning Forum

"We teach the pencil to write“

We at STABILO are currently developing digital writing utensils and applying machine learning methods to teach them how to read. To that end, we evaluate sensor signals from the writing instrument, such as movement data, and attempt to use that to recognize words. For the first time, we’ve been able to use the motion of writing to recognize what was written – recognizing letters and words without external aids, and digitizing the text. In the future, perhaps you will be able to write your text message or WhatsApp message directly with a pen and send it.

One prerequisite for this new technology involves getting sensors integrated into the writing utensils; in other words, embedded systems and embedded microcontrollers have to be small enough for us to actually use them. We also need a communication platform that will work with, say, tablets or smartphones, so we can transmit the data. And the machine learning algorithms have to be so good, and can be programmed so efficiently, that we can execute them on a smartphone. About 15 to 20 years ago, this still cost too much time and money, and it required too much computing power for it to be carried out on these mini-devices. Consequently, it was not possible to have pattern detection or to get a machine learning algorithm to produce usable results.

Machine learning is a new process for designing human-machine interaction to be as comfortable for people as possible: the machine learns from people how to interact with them, and peoplelearn that the machine will deliver the correct results.

"It's about optimization and forecasts"

Repetitive tasks don’t necessarily need a person, they just need a process. That process can then be executed by a person or by a machine. We use machine learning first and foremost to let repetitive tasks be handled by machines.

Data, the right algorithms and a platform those algorithms can run on. The algorithms have to be aligned with the business process; in other words, it is absolutely essential for the algorithm developer to have a solid understanding of the business.

Machine learning is about optimization and forecasts, not about inference, random sampling or populations – that is precisely what differentiates it from statistics.

"We get results faster than with traditional analytics"

I remember arithmetic rulers and slide rules from my father’s time. He was an electrical engineer and, once he’d set up his logarithms and all the rest, they were very good calculators. I didn’t have to learn it at school though, thank God. I never fully understood quite how it worked either. It’s something you have to practice and keep at it to be able to use it effectively. These here, though, these are real computers, you might say. These sorts of machines are not just for private individuals, but often for companies, too. These are the kinds of machines you might expect to crop up in bookkeeping departments, for instance. Just look at the type of keyboards, which vary subtly depending on the application. As far as the mechanical principle is concerned, there are only two possibilities: either the stepped drum or the pinwheel. These are the mechanical mechanisms used to carry the tens and operate the machine. Both principles were invented – and manufactured – by Leibnitz a few hundred years ago. And both were used in such machines well into the 1950s and 60s. It works, so why reinvent the wheel? It’s hard to say which principle was better. That depended on who designed and manufactured the machine. Both principles work perfectly well, and there’s no difference on the outside as far as the user is concerned. Just a box with keys and a crank. The crank is the real heart of the design, though, because this is what you use to enter the numbers into the calculator and perform the operations. Without the crank, the machine won’t do anything.

"Self-learning systems adjust themselves to changing conditions"

Ubiquitous computing is not a new concept: it is some thirty years since Mark Weiser coined the term to describe computer systems of the future that would melt into the background until they became invisible. I believe that machine learning algorithms play a central role here, as they make it possible for technology to take a backseat and to be there only to offer sensible support. I can’t do this with purely traditional data analytics; I also need adaptive and self-learning systems that can adjust themselves to changing environmental conditions or changing requirements. Machine Learning approaches are very well suited for this.

Machine learning involves taking a technically smart system and enabling it to make good decisions that help people, based on a broad set of data.

I think offerings like the Machine Learning Forum are an excellent way to feed expertise about what is possible into companies. What’s more, it’s only when we bring expert knowledge of machine learning approaches together with domain expertise regarding their application that we can come up with genuinely useful solutions. If the industry gathers information about what is possible, then this also boosts its innovative strength. What companies need is a healthy interest in deep learning and machine learning approaches and not to be afraid of them. We’re not doing anything supernatural, we’re simply drawing intelligent conclusions from the data.

Our children’s children will no longer say “I’m going on the internet.”Instead, they will view it as a concept that functions as a backbone, as technology infrastructure that will be a completely normal and integral part of day-to-day life.

To return to the companies – in other words, what they need to contribute – they have to have a healthy interest in and not be shy about collaborating on developments with Fraunhofer or universities. It would also be good if they had their own domain expertise, so they can determine what works and what doesn’t. I often hear “Yeah, we have two terabytes of data, you should be able to do something with that.” That’s not true. It’s not just the amount of data, it’s the quality of that data and the ability to abstract concepts from it. You don’t just need a lot of data, you need lots of individual use cases. You have to develop a certain understanding of them so that you can work with others to create exciting innovations.

“Machine Learning improves our in-house processes "

For Machine Learning, a company needs the right hardware for producing the necessary computing power, but it also needs employees who can apply the entire approach from a technical perspective and data scientists who can do so from a conceptual one. Acceptance among the rest of the employees is important, too – they have to buy into the idea.

Another important point here is that the idea is to make work easier for people, not to replace them.

I’m here today because I want to connect with others who are active in machine learning in the region. I’ve had excellent experiences with these networking activities, including talking with Fraunhofer, academic institutions and industry players.