In 2017, Lee et al. showed that these constructs are equivalent to Gaussian processes in the limit of infinite width. The theoretical foundation (15[1])for the usefulness of these constructs was established by the Universal approximation theorem. Kaiming initialization (15[1])uses a zero-centered Gaussian process with 2 over n variance when constructing these things. The Adam optimizer can be used on these constructs. Max pooling (15[1])can be used to reduce the dimensions (15[1])of feature maps for the (*) convolutional type of these constructs. (10[1])These constructs are trained (10[1])by updating (10[1])their weights (10[1])via backpropagation using gradient descent. Perceptrons are an early version of these constructs that have no hidden layers. For 10 points, name these constructs used in deep learning and AI, modeled on the biological brain. (10[1])■END■