Peter Seeberg is not happy about our choice of title. He is keen to correct a misunderstanding right at the start of the interview: "We talk about 'artificial intelligence', but this is not correct. We are not imitating human intelligence; we are in the business of machine learning – recognizing patterns in data," says Softing's business developer. Seeberg is piloting a "speedboat" at the concern, which helps customers to engage in machine learning – outsourced from their everyday business.
“But a few months from now, I’m sure we’ll be calling it AI, because it just sounds sexier,” laughs Seeberg, and then looks slightly pained. Precise definitions matter to him, as the basis for drawing correct conclusions.
“Data are a new production factor, alongside land, capital and labour, and therefore we have to engage with the issue of data usage.”
As processors get ever faster, they make work easier. Seeberg refers repeatedly to Moore’s Law. He worked at Intel for many years, and must have known Moore personally – he even received an award from his hands. As a reminder: Moore’s Law states that the number of transistors per unit area will double every two years. “Now that is down to 18 months. And my theory is that the quantity of data will likewise double every two years or less.” Knowing this, one can imagine just what might be possible in the near future.
o what changes can we expect to see? “In the past, we saw a problem or an opportunity. The software developer then wrote a code, which was tested, then the algorithm was trialled and fed with data, and a decision was left to the user.” Today, Seeberg thinks, we are headed for a situation where the data come first, are fed to a general algorithm, and the decisions then made autonomously by the system itself.” Seeberg believes that translations will be easier in future.
“All we need is a tiny earpiece, and we can understand Chinese or Croatian.” Self-driving vehicles will also benefit. “Not invented in Silicon Valley, incidentally, but developed 25 years ago in Neubiberg, at the German Armed Forces University. Scientists were able to drive 1,500 km in a self-driving Mercedes. Then the project was shelved. And now the whole subject is back on the agenda in Silicon Valley, and more topical than ever.” Are US firms ahead of their German counterparts? “I don’t think so, necessarily – maybe on the marketing side.”
Because we must not forget that many of the machine learning applications we are familiar with today, including voice control, depend on the LSTM (Long Short Term Memory) algorithm, which was developed by two German scientists from Munich. It was the work done by Sepp Hochreiter (see the article on page 37) and Prof. Jörg Schmidhuber that made intelligent voice assistants such as Siri and Alexa possible in the first place. Isn’t this story a bit reminiscent of the MP3 dilemma? “We’re at the top of our game, and we can achieve great things. We need to remember that. We missed the boat in the consumer market. But now we have a chance to show the world what we can do in machine building.”
And is it true that AI destroys jobs? “Yes and no. AI matters to all of us! No kind of work is going to be unaffected by it. And the more repetitive the work, the more radical the change will be. The worker in industry will produce better-quality work with the aid of AI, while the real work of the radiologist will increasingly be done by AI, because AI can now diagnose patients more accurately than a human clinician.” But why is machine building not really taking off at the moment? “Machine learning is not yet universally understood, and people are unsettled by all the hype; but the technology is not very complicated. Machine learning is based on statistics, and that shouldn’t frighten anyone,” says Seeberg. “Anyone who was good at statistics at university will love it,” he laughs.
But statistics are not enough by themselves. Businesses need to have a vision for machine learning. “We are doing it, not because it is easy, but because it is hard, and because in five years it will be the only way to run a successful business,” insists Seeberg. So what is needed to make it happen? A vision, a strategy –
“the machine builder can detach one department as the speedboat, while the supertanker sails on regardless for now.”
Then there is often a lack of data understanding: where are the data, who is responsible? The production department or the IT department? “Collaboration between departments has to be the aim.
The next step is the hardware. “Businesses won’t need graphics processors to begin with, because we are working with relatively small data quantities and industrial PCs. But this only works if the data contain relevant characteristics, or features, as they are known, in which case we are often processing data volumes in the megabyte range. Terabyte applications are relatively rare. Data training can be done on a standard notebook, the runtime models work on an industrial PC inside the machine – just like a facial recognition model in a modern mobile phone,” explains Seeberg.
And what about the staff? In Softing’s case, they are located in Rumania – the data scientists and data engineers, who ensure the quality of the data and develop the models. There is too much competition for data specialists in the Munich area from big corporations such as Google, BMW and Microsoft. And how does the “learning” work? “Either locally, using open source applications, or in the cloud using AWS, Microsoft or Google.” The final stage is the deployment and regular updating of the models – “then machine learning is up and running, and we can develop new processes, or even new business models.”
To ensure that the transformation of the business and consumer world by AI happens in a reasonably coordinated way, Seeberg argues that every employee must spend at least an hour familiarizing himself or herself with the subject. Those working more closely with the technology need to spend a whole day, or a week, or longer in courses or seminars.