Can pre-registration lead to better reproducibility in ML research?