When Alexa plays suitable music for the occasion as if by chance or suggests the latest book by your favourite author while you are buying a reading lamp - this may be because of Dr. Ralf Herbrich and international development teams. After having studied at universities in Berlin and Cambridge and following various different positions at Microsoft and Facebook, the 44-year-old TU alumnus and graduated computer scientist manages now the team for Machine Learning at Amazon. In respect of Artificial Intelligence, it is the internationally renowned expert’s responsibility to teach computers to see, hear, and speak – and to achieve something that was deemed impossible several years ago: The probability to forecast future events.
Dr. Ralf Herbrich, 20 years ago, while you graduated at the TU Berlin with this subject, Machine Learning was rather considered a marginal issue. What made you to become so enthusiastic about the subject of Artificial Intelligence?
Twenty years ago, research into AI was considered an exotic study with comparably few scientists. I was one of them as the subject interested me very much. I studied, viewing a future that did not exist at the time. At the time I had to dream up all that is happening today. Now we are in a position to carry out processes from the 80’s and 90’s. In the past the biggest challenge was the computer capacity which is available today unlimited thanks to the Cloud. At Amazon more than 1000 people are working at AI worldwide, all of them experts in this field. The field as such is very popular and important for Amazon. This is also something that connects me with Amazon: the DNA, the objective to invent something completely new and to develop something that had once been considered impossible. This is what fascinates me time and again.
Once in an interview you said: “The most exciting thing about machine learning is that science is directly applied here.” How does it currently work at Amazon?
Artificial Intelligence and especially Machine Learning for us at Amazon is an integral part of many products and services. It may be easier to name those products and services at Amazon that can still do without Artificial Intelligence than listing those who work with this technology. For example, Machine Learning enables precise forecasts for requirements: How many red or blue shoes will be required next spring? The real challenge in respect of fashion is - compared to books - that the product range is much larger. Shoes are worn by women and men in different shapes, materials, and sizes. Amazon develops algorithms that can forecast the next season. This is really important to order goods accordingly, have them in stock in a suitable quantity and to be able to deliver on time.
At your Berlin location, a team is also working at speech recognition with Alexa. How does this work?
Alexa is a module to meet the customer‘s convenience. It is, for example, much more convenient to shop by using one’s speech, as customers find it easier to speak freely than to have to use a mouse and keyboard. In short, we help the Alexa team to optimise answers, for example, by an artificial intelligence that can express feelings. The speech modules, so-called phonemes, are put together via a machine learning system such that Alexa can sound cheerful and not like an emotionless computer.
According to a recent survey of the “The Information” magazine, currently only 100,000 customers out of 50 million Alexa-users worldwide buy the smart assistant together with loudspeakers. Most of these want to compare prices and above all want to have a look at the product and touch it. Does this come as a surprise?
No, not at all. It is wrong to think that any decision made by a person can also be made by a machine. This does not work especially when the touching sensation is missing. In Berlin a team is currently working at a project to recognise the degree of ripeness in fruit. It is extremely difficult for a machine to recognise for how long an orange will still keep. As a human being I can use haptics, I can touch the orange, feel, look and smell it to determine the degree of ripeness. A machine that cannot do either of this has simply too few information. For this reason we selected a sensor that can measure light waves the human eye cannot see. This sensor can actually look beneath the orange peel and recognise liquid deposits. This will give enough data to balance the missing touch.
Does Amazon join the orange business in future?
We believe in a larger selection, more convenience and lower prices. During the past 25 years our customers have never asked us to have less selection, longer delivery times or higher prices. It is by far more expensive than necessary to buy fresh goods such as oranges, as half of them perish on their way from the harvest to the kitchen.
Energy efficient algorithms are another subject of the future. What is this about?
We have learning software that wins against human beings when playing chess or ‘Go’. In doing this, however, the algorithms need a hundred to a thousand times more energy than human beings. I run marathons and am naturally aware that I have to save energy, as otherwise there is no energy left when really needed. Currently, the academic research in the section of AI does not concentrate on the energy efficiency of algorithms. However, the more industry uses such forecast calculations, the more important this aspect becomes as the costs for computer capacities will play an increasingly important role in future. At present the human being is still the most energy-efficient intelligence. It will be a long time before computer processors become as efficient as the human brain. It is not AI’s challenge anymore to forecast and become as perceptive as a human being but to be able to use as little energy as a human being.
More than 500 employees are working in the Berlin Development Center of Amazon. An important Machine Learning team from Amazon worldwide also works in the German capital. What makes this location so significant?
For us, Germany, along with other locations such as Barcelona in Spain, and Cambridge in England, is one of the most important locations, due to its scientific strength and highly qualified people. Berlin has three great advantages: Worldwide leading scientists are working at Berlin universities in the sections of Machine Learning and Robotics. There is an unbelievably lively start-up scene attracting many people from everywhere. And Berlin is international. For me it is great and often I find that Amazon colleagues from all over the world also like to work here.
In what way do you benefit from being in the vicinity of Berlin universities such as the TU?
We count on cooperating with many research institutions, among these are universities and research laboratories such as the Max-Planck-Gesellschaft. Unfortunately, companies entice scientists away many a time. I look at this rather critically as these institutions are turning out future scientists. Amazon together with the TU Berlin has set up a post-doctorate model. In this model, post-graduate students work with us for four days and one day at the TU-Institut für Datenbanksysteme und Informationsmanagement [Institute for Database systems and Information Management] of Professor Volker Markl. The Amazon Scholar Programme works in reverse. It is here that scientists can use a holiday semester or a research visit flexibly for projects at Amazon. Currently two directors from the Max-Planck-Gesellschaft are with us. And our CEO sets an example: Jeff Bezos spends four days at Amazon and one day at the space enterprise Blue Origin. The model is obviously working all right.