Cognitive neuroscience has witnessed two decades of rapid growth, thanks in large part to the continued development of fMRI methods. In my talk, I will question what this work has told us about brain function, and will propose some new directions that I see as being crucial to the ultimate success of cognitive neuroscience. First, I will discuss the need for approaches that allow selective associations between mental operations and representations and brain activity. Related to this, I will discuss the need to develop and test formal ontologies of cognitive processes. Finally, I will discuss the need to make research practices in neuroimaging more reproducible.