My Week Exploring AI at Berkeley

I recently had the opportunity to spend a week at UC Berkeley as part of the Lawrence Hall of Science Teen Research Program, where I was in the AI track.

Throughout the week, we explored everything from the fundamentals of artificial intelligence and Python to machine learning, datasets, computer vision, and AI-generated images.

We worked through hands-on projects involving spam detection, social media addiction prediction, voice recognition, CNNs for melanoma detection, image segmentation, and more. What I enjoyed most was getting to move beyond simply using AI tools and actually understand some of what happens underneath them.

For my final project, I worked with a team of five on Speech-Controlled Navigation Using PCA Clustering.

We built a speech recognition system that maps spoken words to movement commands in a maze without using a pretrained speech model. We recorded 180 samples across different words, processed the audio, reduced the data using PCA, and classified each word based on how its samples clustered together.

Our first attempt used “up,” “down,” “left,” and “right,” but it didn't work well. The words were acoustically similar enough that their data points overlapped.

Instead of just accepting the results, we tested six different words and measured how well their clusters separated. We ultimately chose “apple,” “banana,” “water,” and “plum,” which gave us much clearer separation.

We then deployed the project with Gradio on Hugging Face Spaces, allowing anyone to record their own voice and train the classifier to recognize their speech directly in the browser.

The biggest thing I took away from the program was learning how much experimentation is involved in building an AI system.

Sometimes the first approach doesn't work. Sometimes the data isn't good enough. And sometimes the solution is as simple as changing the words you're trying to classify.

I'm really grateful for the experience, the people I got to work with, and the opportunity to spend a week learning AI in a research-focused environment.

It definitely left me wanting to keep building.