Computational Sciences – Machine Learning
Why did you choose to attend Minerva rather than a traditional program?
I believe there are two important and generalizable areas of knowledge that I need to learn throughout my lifetime: understanding humans and understanding systems, which are not mutually exclusive. My previous education gave me a principled approach to understanding systems. I believe that Minerva’s global rotation and international student body enables me to learn about cultures and languages and to build connections within them.
What do you enjoy most about being a part of the Minerva community?
Having culture shock from living with classmates from fifty different cultures. The clash of opinions, habits, and values is challenging but entertaining for the mind.
Tell us about your experience in Hyderabad.
Before Hyderabad, it was interesting to read about emerging markets in a textbook. When I was in Hyderabad, to see emerging markets in person felt different. The city is abound with opportunities for growth and development, and immersing in it gave me new perspectives on economic growth, cultures, and development.
What would you tell another student who is considering Minerva?
Being able to acquire and apply skills around the world in different cities is a valuable experience. A typical Minerva student seeks out opportunities in each city, and Minerva facilitates city immersion and local professional development, as well.
How is Minerva shaping the future, in your own words?
For me, there are two relevant changes. The first is forgoing the traditional campus and using an online platform (the Active Learning Forum) to teach classes. This is a core part of the Minerva philosophy that enables the global rotation: removing the geographical confines of the classroom. The second is the global rotation itself: requiring students to go on a global rotation generalizes learnings across cultures and environments. What is true in one part of the world is perhaps not so in another.
You spent the summer interning at LPixel, a Mistletoe portfolio company in Tokyo, Japan. Describe the project you worked on.
I worked on the research team and focused on the Ischemic Stroke Lesion Segmentation (ISLES) challenge, a medical image segmentation challenge where researchers submit alternative methods for segmenting stroke lesions based on acute CT perfusion scans. An ischemic stroke is a stroke caused by insufficient blood supply in the brain. Plainly put, given CT brain imagery of stroke patients, we were tasked with predicting, to the pixel level, exactly where the brain will be damaged. While a brain imagery obtained from a CT perfusion scan is blurrier, it is significantly faster than a MRI image.
Additionally, I built a server dashboard for LPixel’s R&D department and helped fix their server ventilation problems. The dashboard visualized and monitored LPixel’s graphics cards (GPUs), which are commonly used to compute deep neural networks, and added transparency into the operation. Now, team members are able to see which GPUs are currently being utilized and whether they are fully ventilated, which increases the organization in their research. From the dashboard, I saw that the GPUs were slowing down because of overheating. Simply installing a fan to cool down the GPUs increased the computer performance by approximately 30%.
Why is this research important?
This is important research as an ischemic stroke patient suffers localized brain tissue death each minute their stroke is left untreated. Time is critical and rapid diagnosis and intervention can limit the tissue damage and improve a patient’s prognosis. The clinical rationale for this challenge is to prototype fast, automate algorithms to read CT perfusion scans quickly, and to reduce the time between patient hospital admission and intervention.
What were the different stages of research?
The first two weeks, I allocated my time heavily in anything that would accelerate my research later: developing the server dashboard, resolving the ventilation problems, writing tool modules that took care of data wrangling, experiment configuring, training, evaluating, and visualizing.
After that, it was a process of iteration: reading papers, implementing and evaluating them, and reevaluating what to do next.
What attracted you to intern in Japan?
Coming from China, I have always been curious about the nuanced similarities and differences between the Chinese and Japanese cultures. Additionally, geographically, Japan seemed like a reasonable pursuit for the summer.
How did you learn about this opportunity? Did you engage with the Professional Development Agency to secure this internship?
I attended the Minerva-Mistletoe info session and was excited to learn that there was a machine learning diagnosis and medical imaging research opportunity. I applied because machine learning research opportunities are usually reserved for Ph.D. students and hands-on applied research is valuable for an undergraduate student. The Professional Development Agency had a streamlined application procedure, and it was straightforward to apply. After an application form and several emails, I was matched to LPixel and attended their interview.
Tell me about your career trajectory so far. What steps have you taken while at Minerva to help you pursue your career?
Fundamentally, I believe that business is computable and human societies and inorganic systems will rapidly integrate throughout my life. My long-term goal is to act on this vision. My academic focus has been on data science, machine learning, finance, and marketing. I participate in Civic Projects in each city in the global rotation to gain real-world experience and have been able to apply and develop my skills in my interest areas.
How did your internship at LPixel prepare you for your next steps in your career?
This experience contributed to increasing my global network and further developing my understanding of deep learning and healthcare. In the short term, I have changed the direction of my fourth year Capstone project to expand my research into chest CT scan diagnosis. I know that it is relevant to my long-term goal, but how it is relevant remains to be seen down the road.
What were the result of your research? What is your next step?
At the end of the summer, I wrote an abstract detailing my methodology and uploaded my test results to the ISLES 2018 challenge. As the only non-Ph.D. competitor, my research was not amongst the finalists but placed in the middle, which I am satisfied with but, also, see areas for improvement.
While, the ISLES 2018 challenge has ended, I have come to realize that the bottleneck in healthcare artificial intelligence (AI) is not the model architectures, but dataset limitations. Especially in urgent cases, the imperative is to save the patient, instead of collecting data for a random research. I am excited to continue my work with LPixel over my senior year to develop models on larger datasets in chest CT scans and to investigate how to apply deep convolutional generative adversarial networks (DCGAN) to augment limited datasets.