Amazon scientist Dr. Nashlie Sephus focuses on ensuring accuracy in machine learning
Her work involves many stakeholders and challenges, but it’s an endeavor that is deeply meaningful to Sephus on both a professional and personal level.
Dr. Nashlie Sephus is an applied scientist working on the Amazon Web Services machine learning team. In her role, Sephus is responsible for establishing scientifically rigorous benchmarking and testing methods for machine learning services like Amazon Rekognition, ensuring that the technology produces results that are highly accurate and fair.
This important work involves many stakeholders and challenges, but it’s an endeavor that is deeply meaningful to Sephus on both a professional and personal level.
“I’m a black woman with a southern accent who has been immersed in machine learning technologies as both a scientist and a consumer. I have a personal interest in making sure these technologies are developed and deployed accurately and responsibly. I also want to ensure that the people in my community are in-the-know about the technology and the important topics surrounding it. I’m dedicated to identifying potential blind spots.”
Sephus started her machine learning journey at Mississippi State University, earning a bachelor’s degree in computer engineering, and then studied at Georgia Institute of Technology where she earned her doctorate in electrical and computer engineering. She’s spent more than a decade doing research in the field of artificial intelligence and machine learning. Sephus’ role on the AWS team is, in many ways, a natural progression in her professional career during which she has tackled some of the hardest challenges in computer vision.
I’m a black woman with a southern accent who has been immersed in machine learning technologies as both a scientist and a consumer. I have a personal interest in making sure these technologies are developed and deployed accurately and responsibly.
Prior to joining the AWS team, Sephus was the CTO of PartPic, a service that allowed users to visually search for industrial parts like screws, bolts, nuts, and fasteners simply by pointing their camera at it, thereby allowing them to easily order correct parts from a catalog. PartPic was acquired by Amazon in 2016, and was launched as a feature called PartFinder on the Amazon app. Developing PartFinder required Sephus to tackle non-trivial problems like identifying an object’s orientation in an image, measuring the threads-per-inch of a screw via its image, and developing a comprehensive imaging system for replacement parts.
Now, Sephus is part of a team of scientists and computer vision experts who are driving the fairness and accuracy of Amazon Rekognition on behalf of customers like Marinus Analytics, which uses Amazon Rekognition to fight human trafficking.
“To drive accuracy in facial recognition, we have to ensure that the algorithms are trained on well-balanced, diverse, ethically sourced, and correctly annotated datasets. It’s similar to how using a calculator to solve a math problem doesn’t guarantee you’re going to get the right answer. We want to ensure annotations are accurate and algorithms are explicitly tested for intrinsic biases,” says Sephus.
“We use datasets for benchmarking that are equitably distributed and representative of different ages, genders, and ethnicities, such as Caucasian, black, Indian, Hispanic/Latino, Middle Eastern, East Asian, and Southeast Asian. We are doing further benchmarking on other refined groups with various hair lengths, hair styles, and even further delving into each ethnicity, all with the goal of continuously testing and improving our service on behalf of customers.”
To increase confidence in the results, Sephus and team are working on making sure that testing for Amazon Rekognition and other machine learning products can be measured and replicated. She advocates making the experiments as reproducible as those in typical academic environments.
Sephus is also collaborating with teams across Amazon, leading efforts to gather experts and work on important machine learning topics in a collaborative manner. Some of the activities these groups have tackled include facilitating joint research among research labs and universities as well as organizing a greater presence at external events on behalf of Amazon where Sephus and others across the company present about their work.
In addition to her work at AWS, Sephus is passionate about increasing diversity in the workforce. She founded The Bean Path, a 501(c)(3) non-profit organization based in Jackson, Mississippi whose mission is to sow technical expertise in order to grow networks and communities by providing technical advice and guidance to individuals and small businesses. Programs include hosting tech hours at local libraries, creating engineering and coding programs for youth, and providing scholarships and grants for students or community organizations.
“Having people of different ethnicities, gender, regions, and age groups will help all of us in the industry further improve our technologies,” Sephus says. “One person will raise a valid point that the other person hasn’t even considered, and that helps us make the product better for everyone. “
Sephus is encouraged by the recent positive steps to further diversity across the industry.
“For example, at Amazon we’re doubling down on our presence at conferences that target diverse individuals in technology like Grace Hopper Conference for Women in Computing or AfroTech. I applaud all of these developments. It’s a big part of our customer obsession as Amazonians, and even more importantly, it’s simply right.”