Aarushi Kulkarni

The “T” of ChatGPT: Predictive Models that Discuss, Translate, and Compose

Started July 5, 2023

Abstract or project description

Dive into the fascinating world of Natural Language Processing as we unravel the magic behind cutting-edge language models, like the ones that power ChatGPT. We will focus on the “T” of ChatGPT: the Transformer. We will learn to build and train our own Transformer models, discover the inner workings of attention mechanisms, and explore advanced applications like text generation and sentiment analysis. Learn new ways of thinking about the patterns in language (and even music!) as we explore the potential of AI to generate natural sequences.

In a baseball or softball game, how can one tell the outcome of a pitch based on data?

Started July 8, 2022

Abstract or project description

In a baseball or softball game, the people on defense do their best at each batter to adjust positioning based on what they think the outcome of the specific pitch will be. For example, if the pitch called to throw is a changeup (a slow pitch) the defense will move to the right if it is a right-handed batter, and to the left otherwise, because there is a bigger chance that the batter will hit the slower pitch to that area. However, the defense makes mistakes which costs them in the overall the game. In my personal experience, the defense almost never predicts the event that will follow the pitch correctly. So to solve this problem, my project will be a machine learning model which predicts the event which occurs based on a pitch type, release position, and release speed of the pitch in a baseball scenario. An event would equate to a strikeout, a strike, a ball, a hit, or an out (for hit specifies what type of hit, as in single, double, etc.) As of right now a number based model is what we are aiming for which can be developed into an image based model, which seems more interactive with the user. The goal is to submit the project in a repository on GitHub as well as a science fair with some sort of a smaller write up. The project will use machine learning techniques commonly implemented in artificial intelligence algorithms. Right now the most useful technique seems to be supervised learning because all the data about the pitch and the batter needs to be labeled.