Students Present Real World Data Science Projects

On Thursday, May 21, students from Mr. Pound’s Data Science independent study class presented their real world data science projects to over thirty attendees. 

The class shared their projects with parents, faculty, alumni, and special guests from the tech industry. Alan Steremberg, founder of Weather Underground and mentor to many of the students, played an important role in the design and teaching of this innovative course.

 

Trash Classification

Jessica Le ‘21 and Bruno Coehlo ‘20

Jessica and Bruno used a machine learning model to classify recyclable material in the Commons. The aim of the project is to reduce the amount of misplaced items and cross contamination. The solution included categorizing items into appropriate bins and displaying an image onto a computer monitor screen that indicated the correct recycling bin. The data used to train the model was a combination of primary images and library images. The model results are shown above.

 

Weather Prediction

Thomas Upin ‘21 and Zac Singer ‘21

Thomas and Zac developed a convolutional and residual neural network to predict radar precipitation. They used Google’s UNet paper (https://arxiv.org/pdf/1912.12132.pdf) as a starting point and implemented an extremely accurate solution. Currently, weather forecasting relies on a combination of mathematical modelling and human experience. The use of machine learning algorithms to predict the weather has the potential to revolutionize forecasting and potentially save many lives worldwide. This is a ground-breaking project which would normally be undertaken by a large team of developers.

 

Analyzing Soccer Video

Benji Wu ‘20 and Gavin Taub ‘20

Benji and Gavin used a machine learning model to categorize players, referees and objects on a soccer field and used this data to successfully cluster players by team. They used Facebook’s Detectron 2 library to train the model and then applied KMeans clustering as shown below. The graph below shows the dominant colors for an opposition player. The intention is to use this data to produce possession and pass completion statistics.