Machine Learning Times
Machine Learning Times
EXCLUSIVE HIGHLIGHTS
Effective Machine Learning Needs Leadership — Not AI Hype
 Originally published in BigThink, Feb 12, 2024.  Excerpted from The...
Today’s AI Won’t Radically Transform Society, But It’s Already Reshaping Business
 Originally published in Fast Company, Jan 5, 2024. Eric...
Calculating Customer Potential with Share of Wallet
 No question about it: We, as consumers have our...
A University Curriculum Supplement to Teach a Business Framework for ML Deployment
    In 2023, as a visiting analytics professor...
SHARE THIS:

3 years ago
Wise Practitioner – Predictive Analytics Interview Series: Alexander Wu at Nauto

 

By: Luba Gloukhova, Conference Chair, Deep Learning World

In anticipation of his upcoming presentation at Deep Learning World Livestream, May 24-28, 2021, we asked Alexander Wu, Senior Deep Learning Engineer at Nauto, a few questions about their deployment of predictive analytics. Catch a glimpse of his presentation, Leveraging Driving Data to Gain an Edge, and see what’s in store at the DLW conference.

Q: In your work with deep learning, what do you model (i.e., what is the dependent variable, the behavior or outcome your models predict)?

As an AI-powered driver and fleet safety platform, Nauto models a huge number of risk factors ranging from vehicle state and driver behavior to external road hazards. As a perception engineer, I primarily model the external risk factors, specifically objects like pedestrians, traffic lights, and other vehicles that are likely to impact the safety of the  driver. A huge piece of the puzzle is not only identifying these objects, but also understanding which ones are relevant to the driver at a given time.

Q: How does deep learning deliver value at your organization – what is one specific way in which model outputs actively drive decisions or operations?

Deep learning powers virtually all of our safety features, from realtime, preventative alerts like Forward Collision Warning and distracted driving, to long-term safety features like coaching and risk analysis. With that in mind, we try to leverage deep learning as much as possible when making important decisions. For example, if we’re launching a new feature, the precision and recall of the underlying model will inform us on what thresholds to use across different fleets and vehicles. If we find that the model underperforms in certain settings, we can also tune various knobs to minimize error in production.

Q: Can you describe a quantitative result, such as the performance of your model or the ROI of the model deployment initiative?

Sure. A lot of my work over the last year and a half has gone into optimizing the core object detection model powering our Predictive Collision Alerts. Leveraging some distinctive characteristics of our data, I was able to reduce inference latency on our device by about 30% while simultaneously improving accuracy by over 20%. These numbers may not seem like much at first glance, but even tiny improvements in performance can make a huge difference in the context of driver safety. Technically, the results are also significant because within the field of computer vision, there is generally a trade-off between optimizing inference speed and improving accuracy. Improving both simultaneously is a huge win.

Q: What surprising discovery or insight have you unearthed in your data?

One discovery that really surprised me was just how different our data is from the benchmark datasets that are used to train today’s state-of-the-art models (like COCO and ImageNet). It seems obvious in hindsight, but I think I underestimated the extent of the domain difference, as well as the resulting implications.

For example, while searching for ways to improve our FCW model, I found our external scene datasets to be remarkably uniform. To be sure, there’s still a healthy amount of variance due to mounting position and weather conditions. But compared to a dataset like COCO, our data is highly consistent. This is true across factors ranging from camera angle, object distribution, to overall image composition. Again, in hindsight this is unsurprising, considering that our data comes entirely from devices we built ourselves, mounted in fixed positions on the windshield, facing what is almost always a public roadway. But the resulting implication is huge. It means that all of the recommended deep learning practices – hyperparameter selection, architecture design, and augmentation schemes – were conceived for general purpose datasets with completely different qualities than our own.

Understanding the extent of that domain difference I think gave me a huge push in the right direction towards finding uncompromising optimizations.

Q: What excites you most about the field of deep learning today?

I think there are huge opportunities for deep learning to improve the quality of life for many. Road safety is just one; medical research, disease diagnosis, and fraud detection are all fields where I can see ML having an enormous impact. With that said, I also think there are opportunities for deep learning to be used irresponsibly or even maliciously. And that makes it all the more important to be diligent in how we apply such a powerful tool.

Q: Sneak preview: Please tell us a take-away that you will provide during your talk at Deep Learning World.

Established conventions are widely accepted for a reason: they’re design choices that have proven themselves across a broad range of general applications. But that doesn’t mean you shouldn’t vet a recommended practice before adopting it for yourself. Chances are there’s a significant domain difference between your application and the environment in which that practice was developed. And as long as there’s a domain difference, there are opportunities for optimization.

—————————–

Don’t miss Alexander’s presentation, Leveraging Driving Data to Gain an Edge, at DLW on Monday, May 24, 2021 from 10:20 AM to 11:05 AM. Click here to register for attendance.

By: Luba Gloukhova, Conference Chair, Deep Learning World

Leave a Reply