Deep learning models achieve high performance but often function as black boxes, making their decisions difficult to interpret. The goal of my project is to convert these models into neuro-symbolic systems by extracting human-readable concepts and organizing them into an editable symbolic reasoning layer. This will allow users to understand how the model reaches its decisions and make targeted corrections without retraining the entire network, improving the transparency, reliability, and safety of AI systems for high-stakes applications.

I am participating in SuperUROP because I want to better understand how modern deep learning models work. Although I have built and trained neural networks in my coursework and worked with large language models in research, I am interested in learning how these models reason internally and how their decisions can be made more interpretable. I am excited to apply my machine learning background to research that makes AI systems more transparent and reliable.