Contrasting MacKay's classic Bayesian Occam's Razor with the modern perspective of Wilson et al: a justification for complex, overparameterized models.

Under construction.

Showed that ensembles of sparse subnetworks can vastly outperform dense networks while improving out-of-distribution generalization. This work highlights how sparsity and functional redundancy can be leveraged for more robust learning.

Predictive modeling of NBA shot success using features like shot distance and defender proximity. Trained logistic regression and neural networks. Features analysis by training models with individual variables.