Ralf Herbrich (Schwedt, Germany, 1974) confesses to being a video game lover. In the mind of this computer scientist, with doctorates in programming and statistical theory, that means that he plays with them, but also that he has put his two cents in his improvement. He worked for more than ten years at Microsoft as a researcher, where he participated in driving simulators and in the best-selling Halo 3 for Xbox. Little by little he specialized in systems of machine learning, one of the most powerful areas of artificial intelligence. Then he jumped to Facebook, where he helped close three patents on automatic ranking of ranked searches. He signed with Amazon seven years ago with the mission of directing research in machine learning of the multinational, a task that is carried out in several centers around the world (and which he coordinates from Berlin).
Visiting the offices of Amazon de Barcelona, which houses an R & D team focused on learning natural language, Herbrich slips that he would not mind moving to the Catalan capital. "Barcelona is a fantastic place to live. There are very open people, a lot of movement, scientific excellence, beach … ", he says after pointing to the Collserola mountains, where he ran the morning of the interview.
Amazon started as a book store. What is today?
It is a technological company that tries to focus on a global clientele and that is committed to approaching the client mainly through technology. The massive use of algorithms and mathematical methods to analyze and digitize information is key to selling more easily and cheaply.
What is the function of artificial intelligence (AI) in the company?
It is the technology that allows us to scale existing processes to offer more value to the client. The first books in the Amazon catalog were entered manually. There were people who had to read them and then organize by theme, author, and so on. Now, more than 90% of that process is done by algorithms. There are also cases in which AI literally enables the existence of products. Alexa would be impossible without a voice recognition system supported by machine learning. When I arrived at the company, this technology was beginning to be used in some divisions, such as the purchasing forecasting department. Now I can not think of any that do not use it.
What projects are you working on?
We work in four large areas. One is the automation of model training machine learning, what we call the establishment of super-parameters; the metalearning, or learning to learn. The second area is the translation of voice and text into a language that understands the machine. The third is the automation of the quality evaluation of fresh products: fruits, vegetables, etcetera.
The amount of energy devoted to AlphaGo making a movement is 1,000 times greater than that needed by a human being
How does the latter work?
The 1% of all strawberries, lettuce and others that we buy, we analyze and classify in different levels of freshness. We use infrared, which goes beyond what the human eye can see. These cameras can detect liquids below the surface. We must learn to recognize the appearance of a rotten pear or that is spoiling. It took us four years of work and several failures to get to this (laughs). 50% of all the fruit goes to the garbage before arriving at the refrigerators. We can contribute to lower that proportion. The challenge is to develop this system at an affordable price in order to integrate it massively into our processes.
What else are they working on?
The fourth research group has to do with the forecasts. For example, in fashion: the idea is to be able to anticipate how many units of each garment will be sold 12-14 months in advance, so that it can be produced on time. This forecast is made from trends, models, etcetera. We use deep autoregressive networks, this is, deep learning (deep learning) combined with autoregressive models for time series that allow our developers to foresee their production. As you can see, our way of conceiving research starts with thinking about what we want to achieve and then we do the rest, when the usual process is to advance in different areas of knowledge and then see what comes out.
How do you think the machine learning society in the coming years?
The machine learning it helps us automate repetitive tasks. I think it will change the nature of many jobs, subtracting routine tasks so that we focus on the most creative, just as happened with physical work in the past. What happens with the decision making and the possible biases that the algorithms incorporate? What these do is compress data and repeat patterns, so that if there are biases they will reproduce them. Definitely, in the databases there is underrepresentation by gender, ethnicity, religion … The good thing is that with mathematics we can compensate for these biases, while for an individual it is more difficult to get rid of their prejudices.
Is the potential of the machine learning?
It is true that there are expectations that are not always realistic. All decisions made by the algorithms are based on data collection for years. One of the four that took us to develop the system capable of detecting fresh fruit of which I have spoken is dedicated to collecting data. Then there is the reflection of efficiency. AlphaGo (the program developed by Deep Mind to play Go) is an incredible achievement, but we can not forget that the amount of energy dedicated to the system making a movement is at least 1,000 times greater than that needed by humans. Without greater energy efficiency, it will not be possible to advance much more. General artificial intelligence is a nice theoretical construction, but little else.
Will we get to see Amazon packages delivered by drones?
We are close to it, yes. These devices are increasingly autonomous, thanks also to the AI. What remains to be polished are security issues: we are taking our time to ensure that it is completely secure. But we are in it, do not doubt it.
(tagsToTranslate) ralf herbrich se (t) exaggerate (t) potential (t) machine (t) learning (t) director (t) learning (t) automatic (t) amazon (t) count (t) take (t) four (t) year (t) automate (t) detection (t) fruit (t) fresh (t) fresh (t) intelligence (t) artificial (t) take off (t) review (t) improve (t) efficiency (t) t) energy (t) energy