In this project I use the visualization assistance of variational autoencoders (VAEs) to help understand how regression predictions are made from image data. By jointly training the autoencoder with a convolutional neural network, I can use abstract representations to better understand at a high level what relationships are being learned by the network.
Anthropogenic and Natural Drivers of Groundwater Depletion Across India
Can groundwater loss be predicted solely from knowing levels of irrigation, agriculture, and weather patterns? Evidence suggests it is possible with high degrees of accuracy. Poking around under the hood of our model suggests that irrigation becomes more important over time and weather plays a smaller role.
I investigate using Bayesian machine learning techniques to estimate the effectiveness of time and locally dependent models to predict groundwater levels only through images of the surrounding area. Major takeways: model complexity and interpretability are frequently at odds, and too simple parametric models can do a poor job fitting data if they're too inflexible.
In this project I proved that, under certain conditions, unbiased estimates of true effects can be obtained when we only imperfectly view outcome variables. This situation can be solved by jointly using the information from both proxy outcomes!