––––––– Science Fair 2018/19 –––––––
Playing games is too much work, so for the science fair, I decided to automate the process.
​
Using a genetic algorithm and neural network, the AI slowly learned to play the game at a superhuman level. After a day of training, the bot achieved a max score of 2465, see if you can beat it yourself!
–––––––––– Suess GAN ––––––––––
The goal of this project was to create completely unique characters in the style of Dr. Suess. To achieve this, lots of data was needed; something that wasn't readily available. This meant I had to go through thousands of pages of books and crop each character out by hand.
Once the data was collected, I used the StyleGAN architecture to train the network. Training took around 2 days, but the results were quite promising. Take a look below!
–––––– Science Fair 2019/20 ––––––
For this science fair, I took on the task of automatic text annotation. The idea was to have a simple input device that could find relevant parts of text and summarize the information. Using a simple classification network and layer-wise relevance propagation: within a few minutes, the final product could annotate hundreds of pages with relative accuracy.
​
​
​
​
​
​
​
​
​
​
​
​
​
​
​
A major goal of this project was to make the model as lightweight as possible, meaning it should be able to run on a phone or tablet. This forced me to try new techniques I hadn't used before.
​
https://colab.research.google.com/drive/1x_EA-xCmUr9h7yjim5f2GLlzUe-SvP1s?usp=sharing
​
–––– NPR Student Podcast Challenge ––––
At the end of my junior year, everyone in my English class was tasked with making a podcast. My group decided to look into GPT-2, an AI language model, and its many implications.
​
Sadly, Covid hit while many of us were still wrapping up our projects, meaning we had to shift gears and record remotely. This threw a wrench in some of our plans, but we persevered and were still able to finish the project.
In the end, our group was given an honorable mention in the competition.