As food become more and more scarce in the future, amplifiedat same time by the booming population, there is a great need to analyze anddevelop crops that outperform in specific environments and settings andoutyield to ensure the food security. Corn for example is long used as a mainsource of food for many people. Based on the weather information and soilinformation combined with phenotypic and genotypic information about corn, amodel that will predict yield for different species of corn as well as theirprogenies will have great implication.(Non-disclousure agreement enforced)

Many programmers consider learning and using Deep Learning(DL) these days, given the superior performance of DL in many challenging domains as well as the great demand for DL expertise in industry. However, when developing DL models, programmers, especially DL novices, often have a difficult time choosing appropriate model structures and hyperparameter values. Furthermore, given the rapid development of DL techniques and the rich modality of data, it is even harder for DL experts to keep track of all state-of-the-art model design practices and tricks for different kinds of data. Thus, there is a great need for identifying common model design patterns for different kinds of learning tasks and datasets. In the project, project repositories are mined for common patterns for intuitive hints and helps for designing model architectures.

Project 4
Diffusion Model from scratch, with guided
learning and Moving Average.

Diffusion model uses Markov chain and past prediction technique to recover an image that is completely destroyed by noise. In the paper "Denoising Diffusion Probabilistic Models"(2020), we have seen how diffusion model comes into play and is able to generate identical or similar instances using Markov chain, outperforming GAN in image. I replicated the paper by their formulas and tried to decrease the MSE loss by using Moving Average in Time Series. Using the landscape dataset from kaggle dataset(Landscape Recognition | Image Dataset | 12k Images | Kaggle). In about 300 epochs, I achieved something that looked more like desert, glacier, forest, and so on than the baseline model built from scratch using moving average and guided learning (telling the model what it is looking at.(With improvement and without improvement)
Forest, Desert, coast, mountains, glaciers. (in the same order as follows below)


‍

Schools and degrees

• New Garden Friends School, Greensboro, NC (2016- 2020)
• Purdue University, West Lafayette (2020 - 2023) Bachelor of Science, Data Science

Achivement

• First place in Indiana Dartment of labor competition of employment prediction
• 5th in intermediate level, Purdue Hammer wars competitive programming; 7th in advanced level Hammer wars 2.0 competitive programming
• Honorable mention: ASA Data Fest 2023 Purdue University branch
• Honorable mention : 2020 Hello World Hackathon Purdue University
• First Place in geography knowledge competition Chinese Society of Geography 2016
• 8th in Bank Lending Kaggle competition.