Inside the mind of Amazon’s machine learning innovator, RALF HERBRICH
What is it like to head up machine learning at one of the world’s largest companies? Ralf Herbrich is Director of Machine Learning at Amazon. A seasoned academic, he also led the team at Facebook that uses machine learning to predict users’ actions, and was a Director at Microsoft. I sat down with Ralf to find out why he thinks controlling devices by voice will soon become the norm, and how we will get AI to perform the same tasks as humans without using orders of magnitude more energy.
1. Tell us about yourself, Ralf, and how you came to be who you are today. When you were younger, you wanted to be a fisherman, and I understand this remains a passion of yours?
I am Director of Machine Learning at Amazon with teams in six different locations across the globe. My path started with studying computer science and economics at the Technical University in Berlin before joining Microsoft Research in Cambridge, UK as a postdoc (jointly with Darwin College), and then as a researcher. After 11 years at Microsoft Research, I joined Facebook for one year, moving to Mountain View. Over five years ago, I got the opportunity to start the Machine Learning team at Amazon with a base in Berlin, Germany.
Being out in the nature has always been a source of recharging energy for me. When I was young, I wanted to be a fisherman, but over the past five years, my passion has been running, and you will now often find me running around the nearby Spandau Forest training for a half-marathon or marathon throughout the year.
2. Besides the manifold contributions you’ve made through peer-reviewed papers, you’re also know for multiple co-inventions used by multi-million users, such as Microsoft’s Drivatars and Trueskill ranking and matching system, and many more. It seems like one of your big drivers is that you never accept anything as impossible.
How does this attitude influence your work related to AI today and what’s your secret to successfully proving others wrong?
When you start in a new field, it’s healthy to not question everything and study known results and theories to get a base of knowledge and understanding. However, my learning was always driven by “how can this be used to solve a problem I have?”. So whenever I learned about a new area, I needed an application to make it real. When I heard about Approximate Bayesian inference, I thought to apply it to the problem of skill estimation and matchmaking, and happened to develop the TrueSkill system. Similarly, when learning about game theory and mechanisms, I ended up studying auctions and click-through rate estimation algorithms for online advertisements. A lot of my interest comes from a passion to solve an applied problem. I am currently very excited about learning how we can make machines achieve a human-level intelligent task with human-level energy; this will require me to learn much more about chip design and energy usage for computers, as well as neuroscience.
One little trick I am using to check if something is worth working on is to share my idea with many (academic) friends. If most of them say, “oh yes, this should work”, I am convinced that I am not thinking big enough.
If everyone says “this is impossible” then I will usually re-think my ambition. I typically try to work on ideas where 50% — 80% of my friends are convinced it’s impossible. This has worked well for me so far (of course, that means that 80% of my friends constantly think I am trying to boil the ocean!).
3. We both hold the common view that ML/AI has been changing human behavior for more than a decade already — the way, for example, auto-correction improves our spelling and attitude towards communication, for instance.
From an application point of view, what are the areas of human behavior AI is going to change the most over the coming 2–3 years? And what specific responsibilities do AI practitioners working on such systems therefore have?
I think the two biggest changes AI will make to human behaviour in the coming years is through interaction with machines via voice and gestures. Today, we are very used to interacting with computers through keyboards and touch-interfaces, but AI methods and the availability of lots of labelled data enable computers today to understand and see us as accurately as the human ear and eye. That will change the expectation we have of machines.
I remember my kids once taking a feature phone (without touch) and pressing the screen desperately to advance the pictures and re-dial.
I believe in ten years the standard for consumer devices will be that they are voice enabled like the family of Amazon Echo devices.
I have six different Echos at home: Echoes in the kitchen and bathroom, Echo Dots in the study and living room, an Echo Spot in my bedroom and an Echo Show in my daughter’s room. Interacting through voice with our Echo devices has become second nature, and I think this will happen in many households.
4. At the WorldSummit.AI conference in Amsterdam last October, you brought up a very interesting concept: measuring the efficiency of AI in terms of calories (capturing energy use), as to date, we’ve been relating the quality of AI only to accuracy.
Why is such a measurement necessary and how could it change the way we apply AI?
I believe that this (editor’s note: efficiency) is ultimately the true measurement of AI: if it can achieve a task that requires human-level intelligence with human-level energy, and we are able to deploy it at scale without running out of the only real constant we have in this world: energy! It’s worth remembering that 20% of people’s daily energy consumption is dedicated solely to the brain, so it does take a lot of energy to be intelligent.
Today, the sometimes excessive energy consumption of AI algorithms limits deployment in several ways: in situations where there are large catalogues of items that need to be scored (e.g. millions of products or advertisements for products or billions of web pages), the common practice is to use a low-energy algorithm to reduce the candidate list to a few thousands, and then apply elaborate AI methods on a tiny subset. This is known as an L1/L2 ranking architecture, but it is an artefact of not having energy-efficient ML algorithms. Another use case is consumer electronics: as AI components become basic building blocks of programming (e.g. Amazon, Lex, Comprehend and Translate for object detection, voice understanding, language understanding or machine translation), their energy consumption needs to be controllable to ensure these components are not consuming the battery.
5. At the announcement of Amazon’s AI research centre in Tübingen (Germany) late last year, you stated: “We go where the experts are. And a lot of the AI and machine learning expertise in the world is coming out of Europe.” You and your team are based in Berlin which is a strong statement from one of the world’s leading companies in AI development.
To ensure AI can be applied to large-scale real-world problems, what are the ingredients of a strong and successful team within an organization?
I strongly believe that a team that wants to successfully impact many people’s lives with AI needs to have two properties: it needs to work on scientific problems that are occurring in real customer experience and it does not need to be afraid to work on problems that have no known solution. The latter is what drives every academic, and we have some of the strongest schools for AI research in Europe (for example, the Max-Planck Institute of Intelligent Systems in Tübingen, ETH Zurich, Technical University of Berlin, Cambridge University and University College London, as well as INRIA in Paris).
In order to make sure we are working on scientific challenges that solve customer problems, we use a technique that we call working backwards. We start every research project with a fictitious press release and frequently asked questions (PR/FAQ). Only when we all believe that the experience described in the PR/FAQ is engaging, and we are clear what pain point will be solved for whom, we start the research. This method is somewhat unusual for a research lab but it has served us well over the past five years.
6. You’re also one of the experts in the initiative ‘Partnership on AI’, alongside the likes of Yann LeCun, Eric Horvitz, Francesca Rossi. There, you describe your role as pushing for improvements of customer trust and benefits to society.
In the coming years, what can AI practitioners (data scientists, AI engineers, product developers and business people working with AI) do to help achieve these goals, along with regulators, governments and other institutions such as the UN and Unicef?
The Partnership onAI (PAI) was founded in 2016 and is taking shape as an organisation. PAI aims to study and formulate best practice on the development, testing, and fielding of AI technologies, advance the public’s understanding of AI, to serve as an open platform for discussion and engagement about AI and its influences on people and society, and identify and foster aspirational efforts in AI for socially beneficial purposes. In order to meet these goals, the organisation will host discussions, commission studies, write and distribute reports on critical topics, and seek to develop and share best practices and standards for industry.
The initiative will conduct outreach programmes with the public and across the industry on topics related to advancing better understanding of AI systems, and the potential applications and implications of this technology as they arise. PAI is currently setting up working groups to develop and share AI best practices. The best way to get involved is to let Terah Lyons, the Executive Director of PAI, know that you would like to join one of the work groups.
>>> Check out our other interviews with Joanna Bryson and Ronny Fehling
