Sunny is a student from the United States who attended the Summer STEM Institute (SSI). Since middle school, Sunny has been interested in artificial intelligence and machine learning. In high school, Sunny has participated in his high school’s Math, Science, and Technology Research, where he previously worked on a project using Fourier transformations and recurrent neural networks to transcribe music. More recently, Sunny has become interested in the intersection of machine learning and neuroscience. Over the summer, Sunny worked on a project titled “MRI-based Diagnosis of Alzheimer’s Disease using Deep Learning with CycleGAN for Data Augmentation,” where he created a tool to help physicians diagnose Alzheimer’s disease. You can learn more about Sunny’s project through his research paper here.
Back in middle school, I remember finding YouTube videos and various other online resources about artificial intelligence and neural networks. These resources inspired me to begin working on a few small projects aimed toward learning more about convolutional neural networks. In the beginning of high school, I was accepted into the Math, Science, and Technology (MST) research program at my school, which is a STEM based program where students conduct two research projects a year. Even though these projects weren’t very large-scale, they sparked my interest in research.
Last year, I worked with a couple friends on a project where we transcribed music using neural networks. It turned out to be a pretty difficult project, but it worked out well. We used a Fourier transformation on spectrogram data, converted that into numerical data, and then used labels, which are like notes that would be played on a piano. We trained this on a recurrent neural network so that it would predict which notes were being played at any time.
Back in my freshman year, I worked on a completely different project to maximize the efficiency of dye sensitized solar cells, which are very cost effective counterparts to other types of silicon cells. We experimented with different fabrication strategies and methods for creating the most efficient tool.
I am definitely most interested in machine learning now, mainly due to the most recent research I conducted. Machine learning is such a versatile field with endless applications. Apart from machine learning, another field I’m interested in is neuroscience. A while back, I saw a presentation on Neuralink, one of Elon Musk’s companies. Learning about Neuralink and the work they do piqued my interest in neuroscience, and I believe that there is currently a lot of potential for advancements in neuroscience due to recent leaps in technology. The work Neuralink is doing seems like it’s straight out of the sci-fi movies that I’ve watched when I was younger, and I think it’s incredible many of the ideas that we once thought were far-fetched are now rather close to reality.
The main goal of our project was to create a tool that could help assist physicians in diagnosing Alzheimer's disease. Over the summer, we trained a deep learning model using MRI brain scan image data that we augmented using a generative adversarial network (GAN). Ultimately, the tool we created was able to accurately predict a diagnosis based on images of brain scans.
When I was thinking of research ideas, my initial goal was to find a well known topic and apply a newer technique to potentially improve its results. I just happened to come across the generative adversarial network while I was looking at possible ideas. It stuck out to me because of how realistically it could generate images. With these two ideas in conjunction, I came up with the idea of using generative adversarial networks as a way to augment data. I recognized that this is a common problem in medical data sets because they tend to be really small.
By far, the hardest part of my project was training the generative adversarial network (GAN). I knew it was going to be difficult to begin with because I heard GANs are susceptible to a lot of problems during training. My original plan was to work with the Wasserstein GAN, but then I transitioned to using a different type of network, called the Cycle Consistent GAN, which worked well for the type of project I was doing. Therefore, I ended up with these two separate classes. I realized I could synthesize one class from the other and have the network learn that transformation. This ended up fixing a lot of the original issues I encountered, and this process ended up generating much more realistic results.
On the technical side, I think this project really helped me hone in my Python programming skills. I learned how to work with datasets and use several different libraries. I also learned how to use PyTorch to create neural networks, a tool that I didn’t have exposure to prior to the program. In terms of more personal skills, I’ve also gained a lot of patience because the project was a lot more difficult than I anticipated it to be. Many parts of the project didn’t work out the way I wanted it to, but this pushed me to be creative and maneuver those roadblocks. Finally, I also learned to work well under pressure because the six weeks flew by! The summer was definitely an opportunity for me to improve my time management skills.
I really enjoyed the final Research Symposium at the end of the summer. A couple of students from my lab group were selected to give a presentation, and it was interesting to see how their projects developed over the course of the six weeks. These presentations were really interesting, and I really enjoyed seeing how people made so much progress throughout their projects over the summer.
Working on this project has definitely increased my interest in research overall. Regardless of what field I go into, I'm planning to do some type of research in that field, mostly because of how much I’ve grown to appreciate the originality aspect of research. I also really respect and admire the sense of contribution you gain from the research process, so I definitely intend for research to be a significant part of my future plans. In terms of possible fields, I’m pretty interested in AI and medical imaging, though I don't have a definite path of what I want to do in the future quite yet. I’m hoping to work on an internship centered around bioinformatics or neuroscience in the future.