In this talk, I would like to showcase a few examples of machine learning problems that arise when using brain imaging to understand brain function and its pathologies. I'll first introduce brain images and to derive statistical features. Then I'll discuss how prediction from these images is useful for diagnostic purposes, but also as a windows to understand the brain. I'll highlight specific challenges that arise when learning predictive models from brain maps, and details solutions put forward by our group, namely spatial penalties. Moving beyond well-posed statistical maps, I'll show how a combination of unsupervised modeling and supervised learning can predict phenotypic traits from spontaneous brain activity, recorded without controlling the subjects behavior. Finally, I'll detail how our work builds upon and nourishes a Python software stack that we leverage to interact with practitioners.