Before coming to Clemson in 2006, Melissa Smith spent 12 years as a research associate at the Oak Ridge National Laboratory, a holy grail for research computing run by the Department of Energy and based in East Tennessee. She researched high-energy and nuclear physics instrumentation and began collaborations with the newly formed Future Technologies Group on emerging computing architectures, among other things. Big things.
When she arrived at Clemson, she was searching for a group capable of similar large-scale collaborations that served people in addition to working with people. She has continued to collaborate with some of the top research scientists at Oak Ridge and across the country via high-performance computing, and the results have been prolific.
“I wanted that same kind of opportunity on the academic level,” Smith says. “It was readily apparent to me that this was the same kind of environment Clemson had and that it planned to continue.”
Smith and her group are focused on applying “deep learning” processes, which in computer terms looks like vast amounts of data, to a variety of different problems — so in the case of autonomous vehicles, researching vision and perception capabilities for a car to be able to drive without a driver.
The work is particularly relevant for Smith, whose daughter recently secured her driver’s license.
“When we’re driving a car, it requires vision, hearing, feeling — we’re making controlled decisions,” she says. “Do I brake? Do I accelerate? So in the field of autonomous vehicles, trying in a sense to recreate that dynamic system, it requires people from a lot of different disciplines to come together.”
From a technology standpoint, it requires combining various machine-learning processes. So, machine learning is the process of teaching a computer to carry out a task, rather than programming it to carry out that task step by step. But deep learning amplifies this process to human-level performance, and to do this, computers must be able to handle large amounts of data.
In day-to-day life, deep learning might look like your Alexa virtual assistant understanding what you say and then adapting to your accent, speech patterns and usage to improve its voice recognition technology over time.
For a self-driving car, deep learning looks like developing the vision and perception abilities that actually allow an autonomous vehicle to be autonomous. Fixed cameras use radar (radio waves) and lidar (laser or light) to measure distances and to paint a picture of a vehicle’s surroundings. That video can then be processed in two different ways: one is to identify and classify objects the other is to create a map of the car’s surroundings, something called segmentation.
Every object, at every speed, in every light and weather and traffic scenario, with every road condition must be taken into consideration. Not surprisingly, a giant computer is required to process all that information, which is then used to build the neural network that will ultimately steer a car to drive autonomously.
And then they get to figure out how to downsize it to car scale.
“We think of things in stages,” Smith explains: “Identify problems, come up with a solution and take steps to translate the solution.”