Tony Jebara, the man who’s work helps direct 15% of ALL Internet traffic to Netflix
Tony Jebara is a sabbatical professor at Columbia University and Director of Machine Learning at Netflix. Integral to the successful implementation of machine learning at the company, Tony leads a team driving engagement which has saved Netflix a Billion Dollars yearly. In this interview he sheds light on the development process of AI, the progression of machine learning and more.
1. Who are you Tony, and what influenced you the most to become active in the field you are today?
I am Director of Machine Learning at Netflix and a sabbatical professor at Columbia University. I became passionate about machine learning in the mid 1990s before it was a proven technology. In particular, I started by working on face recognition where many initial approaches used hand-written AI rules to recognize faces. I quickly realized that a data-driven machine learning approach using many real images as training would produce a better system. Ever since, I’ve been heavily involved in the field, publishing, organizing, and attending its key conferences.
2. For over a decade you’ve been working on applying AI to really complex problems and data sets. How would you describe its progress and what to expect the coming years?
Things have been growing at a fierce pace with exponential bursts in progress followed by brief plateaus, which are inevitably followed by new bursts. Rapid progress emerged as the field embraced new ideas like probabilistic modeling, Bayesian inference, support vector machines, convex optimization, learning theory, reinforcement learning and more. The biggest leap was seen a few years ago with the re-emergence of neural networks and novel deep learning techniques such as Generative-Adversarial Networks and Variational AutoEncoders.
These days I expect more progress in causal learning methods as predictive machine learning is limited to finding correlation patterns in the data rather than actionable causal predictions. We need machine learning and deep learning not to only summarize data sets and predict outcomes, but also to show us optimal interventions and policies with which real-world actions can be taken to maximize business outcomes, consumer joy and other benefits.
Causal and deep learning are the next frontier.
3. You think of machine learning as Science 2.0, turning the classic approach on its head by using algorithms to come up with billions of ideas and see which of these are supported by the data. What’s critical to apply this thinking and can it lead to new scientific breakthroughs?
Machine learning flips the scientific method upside down. Instead of starting with a hypothesis and then finding data to support it, you start with a lot of data and you consider billions of hypotheses to see which ones sticks. Of course, there are caveats: you need to be efficient from a computational perspective and from a statistical perspective. Computational efficiency is necessary as considering billions of hypotheses sequentially just takes too much time. So you need fast algorithms that tunnel in on the most promising hypotheses quickly. Statistical efficiency is necessary with many hypotheses as it is easy to over-fit the data and without a rigorous statistical foundation, it is easy to converge onto false conclusions and models which are spurious and are not generalisable.
We should expect new scientific breakthroughs in areas such as precision/personalized medicine where the numbers of variables and the individual deviation are too large to rely on traditional binary A/B testing methods. Through large data and machine learning, we also are making scientific strides in understanding human activity at scale through mechanisms like social networks, recommendation engines and smart city infrastructure.
4. Deep learning is widely used for recommendations, personalization and search. After successfully driving engagement at Netflix saving the company $1B per year, where do you see deep learning’s future but also its limits?
Although the use of deep learning will be increasing due to its continued success in industry, it will not just ‘take over’. Humans will be essential to frame the problems, architect the data sources, identify biases and determine the right business-relevant objectives.
Also, deep learning has fundamental limits: for example it cannot disentangle correlation from causation.
Just because deep learning finds a pattern in the data does not mean that it is an actionable pattern that can be used to create better outcomes. This becomes clear when looking at the correlations humans notice in the world all the time, e.g. “people bring umbrellas to work on rainy days”. Humans know that asking people to keep their umbrellas at home will not stop a rainstorm, but this is only clear as we have a causal understanding of the world rather than a correlational one.
One thing always on our minds is the fact that our data is collected under a certain current system and user interface. These systems bias the data collection process which skews the presentation of the choices to the users. Did the user watch Narcos (one of our Netflix originals) because she wanted to watch it or because our platform recommended and promoted it? Was the cause the user’s innate interest or the influence of the recommendation engine? If we want to train a future engine and a potentially better engine from previous data gathered by an older engine, we need to separate the causal effects from the presentation biases; and this is a new challenge in the arena of machine learning with which researchers are still grappling.
5. You and your team have obviously applied AI on scale serving over 100 millions subscribers at Netflix. Once you said you don’t have one but hundreds of millions of products. What are lessons learned in applying AI at scale and what can others do to leverage this technology best these days?
It is important to remember that AI at scale means helping humans accomplish their goals and improve their decisions. A human is always in the loop, be it in deciding what to watch at home one evening or deciding on a growth strategy for a business. AI is getting better at solving many well-defined technical tasks but it still needs significant progress in modeling and interfacing with humans to help THEM improve their real-world goals.
If AI is like a brilliant jerk that nobody wants to work with, it will not go as far.
6. Leading Columbia University’s Machine Learning Lab with alumni’s such as Adrian Weller, now program director at UK’s Alan Turing Institute, includes supporting the science community, but also AI practitioners via free courses in machine learning and more. How important is it to push the design and development of AI and what can we all do make AI more accessible?
We should continue to teach AI as broadly as possible as it is becoming the lingua franca of our generation.
There is so much demand on all sides. Companies want to hire more and more employees with AI and ML skills. And more and more students are finding AI and ML exciting, fascinating and more accessible than ever before.
Thanks Tony for you time & insight!
