Meet RadiaSoft’s stellar team in our ongoing Q&A series where they tell you about their work, their background, and some of the things that make them who they are.
Today, we speak with Dr. Callie Federer about her current machine learning projects, why she loves Ada Lovelace, and how software development makes science happen faster.
I’m an associate research scientist. When RadiaSoft began to move more into the machine learning (ML) world, they brought me in to work on a number of related projects.
Mainly, I work on the ML parts of various RadiaSoft projects. That includes deep neural network model development as well as the supporting tasks of ML like data cleaning, understanding what data you do and don’t need, and helping my colleagues with the ML portions of their grant proposals.
I got a bachelor’s degree in computer science with a minor in biology. I was unsure about going to industry or grad school, so I planned to take a software engineering job. But I happened to go to an interview for a computational bioscience program in Denver and got really excited about it. I ended up going straight from my undergrad to get my PhD in computational bioscience at CU Anschutz.
“Computational bioscience” is a really big term for why people have interest in using programming/computer science to solve problems in biology. My field was computational neuroscience. I focused on the intersection between machine learning and neuroscience, since neural networks were inspired by how the brain works. I explored the question of how we can use the brain to improve machine learning because the more we understand the brain, the better we can make our algorithms. It goes the other way too. Neural networks are built to be inspired by the brain, so can we use them to ask some questions in neuroscience.
There’s a lot of fear about how AI is going to affect our world. I think what we should be “scared of,” specifically, is misunderstood. The idea that an AI is going to be built to help Grandma get her groceries and then decided to hurt people? Not realistic. When you train your algorithm, all it “knows” is the one thing it’s supposed to do. You tell it, “this is good” and “that is bad” and it works towards good. But on the other hand, people are using machine learning for weapons and wars. The fear should be in what people use machine learning to do, not machine learning itself.
A real danger comes from bias in data. A lot of people come into machine learning and find it’s easy to build their own network. But it’s very difficult to understand what’s really going on inside it. If you don’t have a sense of what’s going on, you can put anything in and get some answers out and not realize the bias in that process.
One example is a big ecommerce company’s machine algorithm that reviewed résumés and decided if a candidate should advance or not. It was in use for a while before someone finally looked at the algorithm and realized it was ranking women’s résumés as lower than men’s.
You like to think that algorithms can’t be biased, but when you put bias data in, bias things come out.
St. Louis, Missouri.
I was a camp counselor in high school. There was a time in my life when I thought I was going to get a parks and rec degree and camp was going to be my life. It’s very hard to describe what it’s like to be a camp counselor—you just become a “camp person.” I had a camp name. It’s a whole alter ego. Someone who wears costumes and performs skits in front of hundreds of kids.
The kids were anywhere between seven and 17. I did friendship bracelets, flag football, soccer, volleyball, windsurfing, and archery.
Ada Lovelace. She was a mathematician and writer born in the early 1800s and she invented algorithms.
At that time people understood the possibilities of machines carrying out basic calculations, but she had the insight that they could do a lot more. She didn’t even realize that in her notes she wrote down the first algorithm. Tragically, she died young, at 37. She did so much in that time. They named a programming language after her called “Ada.” It’s used in some military work these days.
I’ve been working on a project for prostate cancer treatment planning under an NIH grant. When a patient goes into a hospital with prostate cancer, they get a CT scan or an MRI—this project works with CT scan data—and doctors use that data to come up with that they believe to be the ideal radiation therapy dosage. They want to target the prostate the most to kill the cancer, while reducing “off-target” damage to other tissues from the radiation.
We’re trying to normalize that treatment plan. Dosage plans vary from doctor to doctor and from day to day. We take in the CT scans and contour maps of patients’ organs and get a machine learning algorithm to predict what the dosage treatment would have been. (That is, what the doctors said the treatment should be.)
It’s difficult not only from the data perspective, but also because doctors have different opinions on treatments. We’re aiming to have a software program that allows doctors to upload scans and data, and then outputs a full dosage treatment for that patient’s prostate cancer. Essentially, we want to help automate the whole process in order to make treatment more consistent.
I’m pretty good at Dance Dance Revolution. There are definitely people who are a lot better than me, but my talent is sufficient enough that when I go to play it, it will draw crowds. DDR is the perfect meshing of athleticism and nerd, I think.
Conga Party Parrot. I like it because it starts off like Party Parrot, but then does the full conga line. You have to watch for it.
It’s a really good way to show approval and excitement in different situations. A great way to celebrate colleagues’ success is to Conga Parrot them.
Going into industry from academia, I got a lot of shade for not doing “real science.” But RadiaSoft is very much real science. It’s writing proposals, the heartache of proposal rejection, publishing papers, doing high-risk projects—the full gamut. The main difference between what I do now and a postdoc in academia is that we have an amazing software team that makes your science usable much faster than a publication alone ever would.
Personally, I think the combination of publication and software team support is better than academia. The amount of time between writing a traditional publication and, say, having a useful product coming out of that is so long. They run in parallel with my research and make the software for it as I go, like with the NIH project.