Nidhi Parthasarathy is a 10th-grade student from San Jose, California, who participated in Stanford HAI’s AI4All program this June. AI4All is a three-week online experience that introduces young learners to AI through a combination of lectures, hands-on projects, and mentoring. Here, Nidhi shares highlights from her experience in the program.
I have always been interested in computer science, creating apps and coding competitively. But, only recently I started getting interested in artificial intelligence and machine learning. I had listened to an exciting TED talk from Fei-Fei Li about computer vision and AI, and was inspired to learn more. When I got the opportunity to attend Stanford’s AI4All program, I was excited to dig deeper into this area. How exactly do we get machines to “think” on their own? Beyond personal assistants like Apple, Siri, and Google Home, and Netflix and Youtube recommendations, where else could ML be used?
Unpacking the Magic of ML and AI, and Using it Wisely
I joined 31 other enthusiastic incoming 10th graders for this summer program. We learned about the origin and evolution of AI, from the Turing test (“Can a human being interact with a computer without knowing it is a computer?”) to recent developments in ImageNet and its impact on the AI research field. We discussed various areas in AI – computer vision, natural language processing, robotics, and computational biology – and exciting applications in these areas. For example, in the computational biology category, I was amazed to find out that AI models can actually diagnose skin cancer more accurately than humans!
As we studied ML applications, we learned about the difference between learning what to learn and how to learn. We explored different types of ML models and concepts like regression and classification. We learned to measure how well our ML models were performing, beyond accuracy, to other metrics I hadn't been exposed to before, like recall, precision, F1-score, AUC, and ROC. We learned how to use Python for machine learning, to create dataframes and arrays, and graph and analyze data. Step by step, I felt like a magician slowly getting comfortable with performing magic spells.
But just like in magic, we also needed to learn about the cautionary side of using ML. We discussed how AI should be used with careful consideration of ethics and privacy. We learned about “responsible” AI design, to be aware of biases and shortcuts, and the importance of good, diverse datasets. During a presentation by Ariam Mogos from the Stanford d.school, we played a game where we were given a picture of a person and asked to come up with the first health-based topic that it triggered. When we shared our answers, it was fascinating to see how every one of us had different views and definitions of health. The total of all our input together was a very diverse data set that gave a much better definition of health than just one person’s opinion. This activity gave me an actionable way to think about how to address diversity and biases in ML data sets.
Meeting the People of ML/AI
One of the program highlights was listening to lectures from professors, researchers, and industry people about their experiences with ML.
My favorite talk was Stanford HAI and AI4All co-founder Fei-Fei Li’s. She spoke in detail about her personal journey and her research on “making machines see well.” She explained how “to see is to understand,” and highlighted how, “while vision begins in the eye, it truly takes place in the brain,” to illustrate some of the ML challenges she had to solve. Her discussion of the various applications of AI in computer vision to solve diseases was inspiring.
I also enjoyed hearing from Lauren Yang, an autonomous car researcher, about Tesla’s “Smart Summon.” Her talk made me reflect on the responsibility we have in making AI trustworthy enough to drive around without hitting people.
In total, we heard from more than 30 different presenters, including IBM research scientist Stacy Hobson on the future of technology, UCSF professor Marina Sirota on computational biology, and Stanford professor Chelsea Finn on robotics, many sharing great demos of their work and even covering personal and career growth. I wrote a more detailed day-by-day journal that includes summaries of all these talks, which you can read more about here.
Beyond Theory to Practice: Becoming an ML Scientist!
If I had to pick my favorite part of my AI4All experience, it would be the capstone project. Along with three other students, I worked on developing an ML application that analyzed X-ray images to detect the occurrence of disease.
We worked with the “chest” dataset from the MedMNIST database, a multi-label, multi-class dataset where each picture had multiple labels associated with it (one of 14 different diseases). We started by analyzing the dataset and training with a linear model. But this required flattening the images, which we soon realized was computationally too intensive and time-consuming. We instead switched to using a convolutional neural network model (RESNET18). This too required some additional work. We had to convert the grayscale images to color (RGB) and resize them to the optimal size for our model. We then trained and tested our models using PyTorch. Our initial results were a bit of a roller coaster: We initially got terrible accuracy, but then tracked it down to a problem in our code. We then got unbelievably amazing accuracy, but traced it back to another bug in our understanding. And back and forth we went. Finally, after some more work (including some lively team brainstorming and coding), we had a model with a good accuracy (94%) and precision (95%).
We had used machine learning to solve a real-world problem that could change people’s lives! Our learning journey had gone from naive questions to a working ML application in a short time!
We presented our work to other researchers in medical AI and identified opportunities for follow-on research. In particular, while our model has good accuracy and precision, we still have room to improve recall (56%). Using a method called “confusion matrices,” we dug deeper into our data and realized that we had encountered a problem common in the medical world: a significantly higher number of negatives than positives (for example, for one of the diseases, this was almost a factor of 260). This makes sense when we consider that typical medical records have a lot more patients without disease compared to those with disease. This was a great opportunity for me to see in actual action how bias in the dataset could impact the effectiveness of ML applications.
An Inspiration for a Career
Overall, AI4All has been an amazing experience. I learned about AI and its numerous applications, and heard from distinguished researchers in both academia and industry. I also loved the opportunity to work with other students across the country, and particularly appreciated the deep friendships and fun conversations (including our K-Pop discussions 🙂) with my cohort group in our virtual social hour.
I got a deeper appreciation for how research worked in this space, and learned about the nature of real-world datasets and the challenges with extracting meaningful ML insights from such data. And much to my surprise (and satisfaction), I discovered that I love research! I am inspired to continue building on what I learned, to try out new applications, and to continue doing deeper research in this area.
Learn more about the AI4All program at Stanford.
Stanford HAI’s mission is to advance AI research, education, policy and practice to improve the human condition. Learn more.