Inferring Cosmological Parameters with CNN K-Fold Ensembling

This presentation describes an approach achieving 6th place on Phase 1 of this competition while exploring several attempted methods. The submitted approach uses the ensemble of multiple architectures of CNN’s, utilizing the prediction smoothing and benefits of various inductive biases each architecture provides. Preliminary results demonstrated a significant improvement in performance relative to individual models. Ensembling was also found to be the most reliable method in increasing test set performance, significantly out performing more complex methods. Additional methods explored include denoising, different loss functions, and creating synthetic weak lensing maps. This presentation demonstrates the reliability of standard machine learning techniques in the face of dense images, highlighting the significance of noise.