Danny Lange (Denmark) is vice president of AI and machine learning at Unity Technologies. Previously, Lange was head of machine learning at Uber, where he led the efforts to build a highly scalable machine learning platform to support all parts of Uber’s business from the Uber app to self-driving cars.
Danny Lange might not be a name that rings a bell among most leaders, innovators, and entrepreneurs, but it really ought to in the future. Why? Because Lange is the brain behind building the machine learning and artificial intelligence (AI) efforts at Amazon, Uber, and currently Unity Technologies.
Danny was head of machine learning at Uber, where he led an effort to build its machine learning platform. Before, he was the general manager of Amazon Machine Learning—an Amazon Web Services product that offers machine learning as a service.
This impressive and inspiring background made me reach out to Danny to do an interview about his thoughts around the present and future of machine learning and AI—two of the most interesting and disruptive technologies in the world currently.
You used to work as head of machine learning at Uber. How did Uber work with AI and machine learning?
Uber is a business that is entirely data driven—just like Amazon, where I worked previously.
At Uber, they focus on metrics and measurements. I was involved in all three business units at Uber—the core business, the mapping business, and the self-driving car business.
In the core business, we used machine learning to estimate the time of arrival, pairing people up for Uber pool rides and improving the pickup experience by having a computer learn over time where the good pickup spots are in a particular city.
Basically, the core function of the machine learning algorithms is to measure the experience and minimize the friction during a pickup. For instance, in the United States, there are a lot of situations and places in which an Uber vehicle cannot stop. We designed a system that learned this and thus was able to offer a problem-free experience by suggesting both driver and customer meet 20–30 yards away from where the car was booked to stop in the app.
So machine learning was and is all about reducing friction?
Not only, but at Uber it was a big part of the philosophy. In the second business, the mapping business, we used machine learning to build maps for the drivers. We made a system that could read street signs and populate the map. This means that where most companies and people build maps by hind, we used machine learning for the same purpose.
As with the core business, the benefit from doing it like this is that you develop a system that keeps learning and improving. Rather than building a system where you constantly have to update it manually or calculate all possible outcomes, we have a system that learns by itself and continues to improve—without us having to do more than support it a little on the side.
Could a company like Uber have existed before we had machine learning?
It could to some extent, but not at the scale and quality of service we see now.
In the old days, we would have had to base the product on insights and intelligence—built up by people. This could create an okay product but never compete with the current pickup experience, where the system bases its suggestions on millions of rides every day. Just like Uber’s third business unit, the self-driving car business, the idea is to use machine learning to constantly improve the experience. With the self-driving cars, it can be used to detect objects, plan the optimal route, make predictions about bicycles, and so on.
For people that don’t really understand machine learning, what is the idea behind it, and what are the possibilities in the future?
It’s all in the data and in capturing the patterns. That is what the systems are doing.
Let me explain to you how it works. Imagine that you have to build an application that predicts the shipping time for a company. In the old days, you would look at it the following way: There is a place where you pick up the package and a destination address. You then have to build up a complicated set of rules, a rule system, to include the speed of the trucks, planes, delays, and so on. You would try to compute it and maybe end up with two or three hundred rules to try to predict the shipping time.
In machine learning, you don’t think or work like this. Instead, you will base your system on millions of package deliveries that have already been made; this is the most important thing—your data. Within this data you will have the weekdays, the sizes of the packages, how quickly they were delivered, and so on. Within machine learning you call this the ground truth.
So, ground truthing refers to the process of gathering the proper objective, provable data for the test. For example, Bayesian spam filtering is a common example of this. In this system, the algorithm is manually taught the differences between spam and nonspam. This depends on the ground truth of the messages used to train the algorithm; inaccuracies in the ground truth will correlate to inaccuracies in the resulting spam/nonspam verdicts.
In the case of the shipping, you will then have millions of packages delivered, and the computer can learn a statistical model. When you feed in a new delivery, the system will use the statistics to predict the shipping time based on history.
What we have learned is that this system will always outperform the rule-based system. We have stopped trying to understand the rules. Instead, we leave it to the machine learning system to do that.
Instead of having to maintain the rules manually, you leave it to the machine learning algorithm to do so. Correct?
Yes, and since the world is constantly changing, this will improve the predictions a lot and save a lot of manpower and time. In the case of machine learning, you can monitor the feedback and constantly measure how good your model is.
Is that the same as AI, which is another hot technology and trend at the moment?
To me, artificial intelligence is about how a system is being perceived and how a system presents itself.
If you look at predicting shipping time, that is not really intelligence. But when you start taking an entire organization and have everything it does—from predicting shipping times to detecting hazard materials with computer vision, self-driving trucks, dynamic prizing based on demand, and so on—it actually appears pretty smart and intelligent to me.
That was our philosophy at Amazon: that the whole company would start appearing more intelligent to the customer. And at one point, we could actually claim that we were a really smart organization.
We’ve been talking for AI for a long time, but it hasn’t really drastically changed industries and business models yet. When will this happen, do you think?
That is a huge shift underway. Of course, not everyone can be Amazon, but we can all learn from them.
In the case of Amazon, their mindset is that they need to be able to beat every retailer out there. They do this by knowing you better. Getting you things faster. Giving you more reasonable prices. Offering you more than a billion products. And all of this is only possible by integrating machine learning and AI into every aspect of the business and business model.
Uber is the same. They can run out of San Francisco but beat a taxi company in almost every country. Well, at least where they are legal. Both companies are using the technology to a scale that has never been seen before. It enables them to run a service in a far away country better than the people actually living in the country.
We’re seeing a development where you will be in trouble 24–36 months from now if you don’t start taking machine learning seriously. It will happen especially in industries such as transportation, shipping, finance, and retail, but all kinds of companies and leaders should look into this much deeper.
Of course, the big companies have an advantage due to the amount of data they often have. The startups lack this, and data is increasingly becoming king.
For example, you may be able to build a better app with a better backend than Uber, and pay a crew of drivers more money, but if you don’t have the data to deliver a consistently better pickup experience, all of that might not matter at all.
So you have to be Uber or Amazon to succeed?
Fortunately not, but you have to start collecting data and working with machine learning.
Imagine that you’re a company building houses. There will be many examples—a shipping company without shipping data. A lot of startups are running into the problem that they don’t have the data. Currently, we build homes and offices based on the architect’s creativity and our history and experience of building houses. However, in the future, we could use data and AI to totally change the way we think about the design of houses.
At an artificial space like Unity, where I work with gaming universes, you could build and simulate a family living in that house. You can build virtual characters to live in the house. You can accelerate and have thousands of families living in the house virtually for a year but calculate the results in weeks.
You can then ask your system to figure out how the experience was. Is the hallway too narrow? Was there enough room for furniture? When we sit down and watch TV, where do we prefer the TV to be placed?
By having virtual people living in these homes, we can optimize home building for emotional optimization.
Really? Virtual people are testing products and services before they are launched?
Yes. And in numbers that would never be possible in the real world and without the risk of actually building the wrong product.
It sounds a bit like crazy magic!
It’s not. It’s just applied machine learning. It’s coming to us now because of the development in computing power.
If you look at the growth in computer power, it’s currently going through the roof. In the old days, smart people would try to create such models in their head and implement them in highly efficient programming languages.
They would be highly efficient implementations, but always approximations. This paradigm has changed.
And in addition to this, growth in computer power data is growing rapidly too, right?
Indeed. You have to remember that it’s only 10 years ago that the first iPhone was released.
Before then, nobody captured your location. Now everything captures your location and everything else.
This means that today we have all this data.
You also talk about the concept of reinforcement learning, where data is also vital.
Yes, and reinforcement learning is actually what I spend most of my time on. In traditional machine learning, you work with the concept of the ground truth. In reinforcement learning, you let the system learn the truth itself.
You provide the system with some very fundamental guidelines. We call those rewards.
We then provide the system with some ethics. Tell it what is good and bad. For example, getting run over by a car is bad, and getting to a destination in time is good. That is it—a very fundamental and low-level set of rules.
The system will now experiment its way toward perfection. The key thing is that there is no algorithm dictating traffic law. The system will automatically aim to not kill people and get there in time.
These systems can learn very complex things. If we talk about poker, they can learn to start bluffing—learn to fool you, make some moves that will basically hide the real intention. So these systems build strategic behavior and do things that work—at a very high rate, figure out these subtle strategies.
Going back to the Amazon and Uber cases, these companies might have all that data, but a normal retail store does not?
Correct, but I envision a development where cloud services will be made available to everybody, and in these services, things like face recognition and machine learning will be included.
This means that a physical retail shop can use that cloud service to recognize and track every customer coming into their store, the moment you step into their chain of clothing stores. If they can recognize your face, the employee will know your name, what you looked at last time, what you prefer, and so on.
Staff will be able to give you personal assistance, and I think that this development may actually change the dynamics once again and might revolutionize the space of retail. So machine learning and AI is fundamentally changing the way that things are working around us.
What is your advice to leaders and companies reading this? How should they navigate this field in the future?
It’s about culture more than projects. Stop creating projects and test projects. Instead, think about how machine learning and AI can fundamentally change how you are working and innovating.
Also try looking at other industries and what you can learn from them. For example, at my current job at Unity Technologies, our gaming world is actually much more than just gaming.
What we work with is real simulations of an artificial world. When you play a game, it doesn’t really exist in reality.
You can use the same methodology in other industries. If you’re building a self-driving vehicle, it can be very dangerous to drive around on real streets to test the vehicle. Instead, you could create a virtual street scenario and train the vehicles in this. You put in pedestrians, parked cars, intersections and so on, and the machine learning algorithm in the car will then in fact be able to train real-world action in a virtual space.
So I would advise most companies to change their mindset and become truly data driven.
IT IS A CAPITAL MISTAKE TO THEORIZE BEFORE ONE HAS DATA.
Sherlock Holmes