Meet the Researcher: Anna Choromanska, Optimizing Deep Learning Models for Autonomous Vehicles and Robotics

‘Meet the Researcher’ is a new series in which we spotlight different researchers in academia who are using GPUs to accelerate their work. In this first edition, we spotlight Anna Choromanska, an assistant professor at the New York University (NYU) Tandon School of Engineering and a recipient of the Alfred. P. Sloan Fellowship.

Choromanska’s NYU lab focuses on deep learning optimization and machine learning for robotics and autonomy. Her latest research in partnership with NVIDIA’s New Jersey Autonomous Vehicles lab, Multi-modal Experts Network for Autonomous Driving, was accepted at this year’s IEEE International Conference on Robotics and Automation.

She won the Alfred. P. Sloan Fellowship with her 2019 research paper, Leader Stochastic Gradient Descent for Distributed Training of Deep Learning Models, in the Neural Information Processing Systems Conference (NeurIPS), 2019.

The excerpt below is a summary of a conversation with the NVIDIA team and Professor Choromanska. Minor changes have been made for brevity.

What is your research focused on?

I really have two main research areas of focus, one is deep learning optimization work on multiple GPUs, and the other is the autonomous driving project. 

Deep learning optimization relates to designing better deep learning algorithms. This is a challenging problem to solve because a certain function has to be optimized to train deep learning models and this function is not very well behaved. Specifically, we don’t understand many properties of this function. So, I’m interested in studying the properties of this function and consequently designing algorithms that adapt to these properties and can efficiently optimize this function. 

I also do a lot of work in different aspects of autonomous driving with NVIDIA, starting with building algorithms that help you debug the system. For example, we are trying to detect whether a neural network used in an autonomous vehicle is picking up on visual cues that are reasonable for the network to pay attention to while driving. This is very useful because you can see whether the network actually sees the visual cues that it should pay attention to in order to safely steer the vehicle. 

How does your collaboration with industry partners benefit your students? 

I have over 50 NVIDIA GPUs in my lab. Having the latest tools in my robotics and autonomous vehicles helps students with the constant flow of ideas, it spurs discussions, and the students are more engaged. A lot of my students have also gone on to be interns at NVIDIA.

I think it’s really valuable for students to see the entire process of research from the very beginning, discussing what ideas to implement, proof check, how to scale up the problem, and how the solution is pushed to the point of deployment. If you don’t have collaboration with the industry, it just becomes an exercise of publishing research. 

How are GPUs important to your work? 

Without GPUs, there would be no research of parallelizing optimizers for deep learning, without GPUs there would be no research in my lab. There would be no way to conduct research.

The availability of NVIDIA hardware and software makes it possible to scale and enables the pursuit of research directions which otherwise would be out of reach. 

Choromanska at MLconf in 2019

Learn more>>