NeuralAttentiveCircuits

General purpose models typically makefewassumptions about theunderlyingdata-structure and are known to perform well in the large-data regime. At the same time, there hasbeen growing interest inmodular neural architectures that represent thedata using sparsely interacting modules.