Machine learning models, particularly deep artificial neural networks (ANNs), are becoming increasingly influential in modern neuroscience, often being employed as models of cognitive and perceptual systems. However, these models are often complex and opaque, leading some to worry that, by utilizing ANNs, neuroscientists are trading one black box for another. On this view, despite increased predictive power, ANNs effectively hinder our scientific understanding of the brain. We think these worries are unfounded. While ANNs are difficult to understand, there is no fundamental trade-off between the predictive success of a model and how much understanding it can confer. Thus, utilizing complex computational models in neuroscience will not generally inhibit our ability to understand the (human) brain. Rather, we believe, deep learning is best conceived as offering a novel and unique epistemic perspective for neuroscience. As such, it affords insights into the operation of complex systems that are otherwise unavailable. Integrating these insights with those generated by traditional neuroscience methodologies bears the potential to propel the field forward.