Senior Director Of Neuroscience
United States
Computational neuroscience focuses on the detailed implementation of computation, studying neural codes, dynamics and circuits. Artificial neural networks tend to avoid precisely designed codes, dynamics or circuits, favoring strong optimization of cost functions (violent search), usually using simple and relatively homogeneous initial architectures, using structured architectures, including for attention mechanisms, recursion and various forms of short- and long-term memory Second, the cost function and the training process become more efficient. Second, cost functions and training processes become more complex and evolve over time. Here we think about the brain in light of these ideas. We assume that (1) the brain optimizes the cost function, (2) the cost function is diverse and different at different locations in the brain at different stages of development, and (3) the optimization operation is performed within a framework that is pre-structured by behavior and matches the corresponding computational problem. In support of these hypotheses, we argue that a series of implementations of Credit Assignment via multilayer neurons is compatible with our current knowledge of neural circuits and that some specialized systems of the brain can be interpreted to achieve efficient optimization for specific problems. Through a series of interacting cost functions, such non-uniformly optimized systems make the learning process data efficient and precisely tailored to the needs of the organism.
