Adam Marblestone, Greg Wayne, and Konrad Kording published “Toward an Integration of Deep Learning and Neuroscience” in Frontiers in Computational Neuroscience in September 2016 (Vol. 10, article 94). It is one of the most cited attempts to lay out a research agenda joining the two fields.
The paper advances three hypotheses. First, that the brain optimizes cost functions, much as deep networks are trained to minimize a loss. Second, that these cost functions are diverse and can differ across brain regions and across development, rather than being one global objective. Third, that the brain’s architecture is specialized to make this optimization work, with structured circuits and pre-wired components supporting the learning rather than starting from a blank slate.
The authors used this frame to reinterpret a wide range of neuroscience findings and to argue that the apparent gap between detailed biology and abstract machine learning could be bridged by asking, for any neural circuit, what cost function it might be optimizing and by what mechanism. The paper also took seriously the credit-assignment problem and the question of whether the brain approximates gradient-based learning.
For a general reader, this review is a readable map of why so many researchers came to believe that deep learning is not just a tool inspired by the brain but a genuine lens for understanding it, an argument that has shaped the NeuroAI field ever since.