We caught up recently with UBC Computer Science’s Associate Professor Frank Wood to learn about his research and spin-off company, Inverted AI, which drives impact from Artificial Intelligence (AI) research by providing the autonomous-vehicle (AV) industry with advanced simulation technology.
What started you on your career path? And how did you get interested in AI?
In my youth, I wanted to figure out how the world worked, and started down the Physics path. I ended up in Computer Science motivated to build a machine to understand exactly that.
Along the way I became quite curious about how the human brain works. After studying Computer Science at Cornell University, I ran a couple of successful start-ups. After six years of entrepreneurship, I went back to grad school at Brown University where I continued to study Computer Science with a concentration in Computational Neuroscience. Then I did a post-doctoral fellowship at Gatsby Computational Neuroscience Unit. Next, I did a professorship at Columbia University for Statistics and another at Oxford University for Information Engineering; I successfully exited a couple companies along the way and have been focused on the mathematical foundations of methods and algorithms for AI ever since.
I came to UBC in 2018 where I could be both an academic in Computer Science and an entrepreneur, deploying my research for immediate, real-world benefits.
What drives you in your work?
As an academic, I’m curious about the computational side of the brain — how electrical stimulus turns to sensory perception, and how to engineer intelligence. My research interests include predicting human behaviour using artificial intelligence.
As an entrepreneur, I’m interested in the multi-trillion-dollar question of creating a non-human agent capable of performing tasks that humans do. An agent can be thought of as a robot or, more accurately, as a computer program that directs useful actions, for instance, generating steering and acceleration controls for a car.
With colleagues, I lobby for a definition of such an agent that is arguably more refined than the Turing test—the go-to method for determining whether or not a computer is capable of human-like thinking. It starts with Mechanical Turk, an Amazon product that cost-effectively and efficiently farms “hard” tasks out to on-demand, human workers, tasks like labelling images with whether-or-not stop lights are present for example. As intelligent agents get “smarter,” they are capable of doing more and more tasks that in the past required a human. My definition of an intelligent agent states that such an agent must be able to not only make money posing as a Mechanical Turker, but, more importantly, to be able to continue to make money over time as labor rates are driven to zero on tasks that become automatable.
I am motivated by seeing the immediate application of my AI research come to life in the company I co-founded with Dr. Adam Scibior, Inverted AI. The core of our work and research is the development of agents that act like humans while simulating driving, a task that would otherwise require highly skilled and expensive human labor.
Tell me about your research.
My work stems from the area of deep generative modeling and probabilistic programming — using statistical modeling to develop artificial general intelligence that behaves like humans. Specifically, I focus on using methodology that embraces uncertainty, as humans exhibit a wide, seemingly random distribution of behaviors.
What problems are you solving?
As noted, it’s exciting for me to immediately apply my research to real-world challenges, like I am doing with Inverted AI. The venture is about to launch a cloud-provisioned API (application programming interface) for controlling non-playable characters — agents — in simulation that are reactive, realistic, and behaviorally diverse. Essentially, the agents act just like humans.
The target market for the company is the autonomous vehicle industry. The AV industry tests and trains vehicles using simulation continuously, and the automated non-player characters used at all levels of testing and training are crucial. A Rand study quotes 16,000,000,000 kms of driving is required to fully verify everysingle release of an autonomous vehicle software system. So, the industry challenge at hand is logging the extensive distance and the amount of time that takes in order for the required proof of safety in a realistic environment.
Clearly simulation is the only answer.
What is your proposed solution?
Inverted AI’s predictive human behavioral models and derived from video data. We use proprietary techniques to transfer these models into simulated environments where we use them to drive human-like non-playable characters (NPCs) that are statistically accurate and very similar to humans in terms of reactivity, realism, and diversity of behaviour. And most importantly, the agents (the NPCs) match the distribution of human behaviours of people in the real world.
What is the benefit?
The benefit of Inverted AI is that it dramatically decreases the time and money required for autonomous vehicle validation and training using artificial-intelligence simulation technology. Ultimately, Inverted AI can lead to safer AVs, more rapid training, and quicker production.
Tell me about your experience as an entrepreneur at UBC.
It has been my dream to be both a professor and an entrepreneur. I am very grateful to have landed in a place that enables both. In the past, UBC has had a reputation as being difficult to work with from the entrepreneur/investor perspective. I am delighted to say, thanks to their flexibility in working with me on Inverted AI, UBC is evolving to be more generous towards spinoffs and start-up companies. It’s a very welcome change, one I’m pleased to help bring about.
If I were interested in studying with you as a graduate student or post-doctoral fellow, what suggestions would you have for me?
I want to work with intelligent, thoughtful people who can leverage my work in future applications to advance the field.An excellent strategy is to propose research that extends and explicitly cites my recent work.Postdocs with strong programming languages, statistics, and applied machine-learning skills are also welcome additions to my team. Visit my department webpage for specifics. I utilize the Mitacs program to help support the best talent available. Please review the model to make sure you’re eligible for funding through this program.
How can I tap into your research for business solutions?
Predicting human behaviour using artificial intelligence has many applications in gaming, robotics, city planning, etc, and I’m interested in helping advance industry writ large using this technology. As noted for students above, I utilize the Mitacs program to cost-effectively source and support next-generation talent in developing solutions to real-world challenges. If you have an application in mind, please be in touch via my department webpage.
What’s next on the horizon?
My UBC graduate students and I publish very regularly in leading AI/ML venues. So, there’s lots of leading-edge research underway.
In one, we have discovered a new way to use a critic — a sort of a back-seat driver — to help us reliably find human-like, infraction-free solutions to difficult driving situations. This algorithm allows us to create behaviour that is nearly infraction free like humans. It’s getting close to being perfect. This has extremely powerful applications ranging from better AV simulations to longer and more coherent text generation.
Simply put, the potential is monumental.
Dr. Frank Wood and Dr. Berend Zwartsenberg, former PLAI group postdoc and Inverted AI CSO, discussing research; in particular a new probabilistic model for generating initial conditions for autonomous vehicle simulation.