Ralf Herbrich is one of the three masterminds behind Amazon’s artificial intelligence, the secret formula responsible for the exponential growth of the company that has revolutionised the world of commerce.
He argues that ‘smart machines’ will create jobs and that designing algorithms that respect the privacy of users will be a crucial research field for the future.
Ralf Herbrich is director of Machine Learning Science at Amazon. Based in Berlin, he is in charge of AI together with two colleagues based in the US and India. Prior to joining Amazon, he worked at Facebook leading the team for building machine learning infrastructure to predict users’ actions. He also worked for Microsoft as director of Future Social Experiences (FUSE) Lab UK.
Herbrich spoke with EURACTIV’s Jorge Valero.
Amazon is arguably the fastest growing company on earth. Can the firm sustain this growth rate over the next five to ten years? What will the driving forces be?
The short answer is yes. We have to stay focused on the three ingredients to excel in customer experience. If we keep focusing on increasing the selection on digital and physical products; if we stay focus on lowering the prices constantly; and if we stay focus on convenience, in retail for example by helping you to find faster what you want and shipping it to you faster, then I believe we can keep this growth.
Sometimes it is not enough to remain loyal to your core principles, you need to try new ways of implementing them because competitors copy your strategy. What will your future growth engine be? What role will artificial intelligence play?
This model is only capable of scaling with the help of algorithms. Today, a large fraction of our products, 90% in some cases, is forecasted by algorithms. But still, there are some people hard to forecast by algorithms. For example, when we look at X-Ray expertise, many books and annotations are made by algorithms, because they understand the structure of language, but not 100% because there are still some language constraints. The three principles will not scale up by manual labour only. They keep accelerating and growing only if we augment and supplement human intelligence with machine intelligence.
Machines need data to learn. Data comes mainly from two sources: people and sensors. Increasingly the latter. This creates issues such as data ownership and privacy. Do you think these aspects will slow down the development of machine learning? How do you address citizen concerns regarding data privacy?
The question of data governance is important. That is why we never release any of this data. But data also belongs to our customers. That is why if they want to store it in Germany, we keep it in Germany. We don’t move it. Data from the public is publicly usable, and we make use of it. There is another interesting science question: how you can design learning algorithms that keep the anonymity of the individual. The question of privacy and machine learning is a very active research area. No one in the world has the full answer to how you can make an algorithm aware of privacy, but it will be an important one.
As you know there is an ongoing debate, especially in Europe, about the free flow of data. The fact that you favour data storage locally, is it because you are German? Do you have clear differences when you discuss the flow of data with your colleagues = based in the US or in India?
We have the same approach. Internally we also use Amazon Web Service. When we store data in the data centre it is not going to be moved. For example, the Japanese team stores data in Tokyo, or the American team in Virgina, or the German team in Frankfort. And this is importance because of the data governance I mentioned.
Nowadays, machines are not only learning by themselves but also teaching to each other, as the MIT has proven. Given the fast pace of progress in artificial intelligence, and some risks associated to how far it could go, do you think we need to install a kill switch in the robots?
The deployment of algorithms that improve over time is becoming more important. Amazon has formed a partnership with Microsoft, Facebook, Apple, IBM and Google for the safe deployment of AI in society. I don’t think we need a kill switch but we need good practices. You will need such a switch if algorithms develop some sort of conscience. But the algorithms we are talking about extract patterns, they recognise similarities.
Besides this group, should there be some additional regulation to avoid the unintentional consequences of AI?
The group is not about self-regulation but about sharing best practices, and we also include the best people from the academia, to steer research on the impact of AI on society.
Do you think robots and artificial intelligence will destroy jobs?
I don’t think there will be job losses, there will be job changes. It is a typical fear because the nature of the jobs changes with technology. Jobs that don’t exist today will be created. I will give you an example. When we built the machine translation system, we realised that we still need people that collect the data to decode the language, to fine-tune the parameters of the system. The scientists weren’t the right people to do it. So we have to create a new job called the localisation engineer. We now have employed more localisation engineers in Berlin than researchers that build the system. That is a job that does not exist without a machine translation system. For me, that is an indication that new jobs will be created.
As a customer, wouldn’t you be concerned if a company like Amazon has so much power in so many fields?
I would be concerned if the company was focusing on shareholder value, on financial incentives. What I find quite reassuring is that Amazon is a company that recognises customer come first, and the trust of them is the highest currency. The engagement is what matters.