The most recent Thrifty Food Plan developed by the USDA in 2021 serves as the basis for the maximum Supplemental Nutrition Assistance Program benefit allotments. The underlying mathematical optimization model produces the recommended food plan for the lowest income population in the country that balances nutrition, cost, and minimal deviation from people’s current consumption patterns. To address food-induced carbon emissions, we add carbon emissions analysis to the optimization model for the Thrifty Food Plan 2021. Synthesizing food-induced carbon footprint data from DataFRIENDS, we use machine learning, specifically the DBSCAN algorithm, to detect footprint clusters in food categories, and select them for higher and lower carbon subdivisions. With carbon footprint subdivision of selected food categories, our model produces a weekly food plan for a family of four that reduces carbon footprint by 11.3% (from 106.2 kg CO2equivalents to 94.2 kg CO2equivalents) while meeting all the requirements on nutrition, budget, and resemblance to individuals’ observed diets. Our work shows promise in applying machine learning to the increasingly complex and interdisciplinary data in food and environmental sciences. Our model demonstrates a way to develop a sustainable, affordable, nutrition-sufficient, and culturally acceptable model for the country's lowest-income population. Given that the Supplemental Nutrition Assistance Program feeds one out of eight Americans, our work represents a meaningful contribution to help fight climate change and dieted-related chronic diseases.

Video action recognition is a critical challenge in a wide range of practical applications. Most of the video recognition models are computationally expensive and energy intensive. EfficientViT is a new family of vision transformers that offer high efficiency while maintaining state of the art accuracy on vision tasks such as image classification and semantic segmentation including on daily devices and cloud computing. However, the effectiveness has only been validated on images so far. In this project, we extend EfficientViT to video processing by replacing 2D convolution with 3D convolution. The updated EfficientViT model B1-r288 was successfully trained with the Epic-Kitchens-100 video dataset, achieving top 1% verb accuracy of 27.7%, a 5.4% improvement than the original model.