Collaborative Projects

Post-hurricane damage assessment is one of the most crucial steps of disaster response. To mitigate the cataclysmic aftermath of a hurricane, we propose a robust auto-encoder model that classifies satellite images of building images into ‘unaffected’ and ‘flooded’ regions. We classify images of buildings that were collected post-Hurricane Iota, to achieve a test accuracy of 98%. The quick access to geospatial data enables us to automate damage assessment and, ultimately, expedite rescue operations by keeping manual site assessment to a minimum. 

Intending to understand the effectiveness of COVID-19 containment policies adopted across different geographies, we analyze COVID-19 trends in India, South Korea, and Italy. We propose an ML-based strategy that uses the previous analysis to predict which policy- early lockdown, late lockdown, or closed contact tracing- will work well for Brazil’s population. This open-ended problem required us to define performance metrics indicating the effectiveness of these containment policies based on the actual deaths, recoveries, new cases, and population trends. According to the projections made by our ML model, we infer that South Korea’s close contact tracing holds the promise of reducing the COVID fatality rate in Brazil. 

We design and deploy a Neural Collaborative Filtering model for personalized restaurant recommendations in Toronto. The features are derived from Yelp’s data set, which comprises information on businesses, business ratings, user preferences, restaurant reviews, and customer tips. Our model has a test loss of 1.4, and the predicted ratings are close to the actual ratings within the rating range of 3.5-4.5. 

We use the ATLAS dataset to identify the Higgs Boson. Our aim is to classify particle collision events into those that produce the exotic Higgs Boson particle (signal events) and those that do not (background events). We participated in Kaggle's Higgs Boson Machine Learning Challenge (2014), and our LSTM model stands out with an AMS score of 2.35. We also compare the performance of several ensemble-based models against customized RNNs.