(a) [10 pts, page 7] State the Universal Approximate theorem of feed forward neural networks (b) [20 pts, page 7] Draw a sparse auto-encoder of 1 hidden layer with 2 hidden nodes where input and output have dimension 5. Suppose this network is initialized with 0 for all its weights. Using back propagation algorithm, show how these initial weights will be updated with training data x = (1, 0, 0, 0, 0). Use the sigmoid function as activation function and 12 loss function. Assume no bias terms. (1,0,0,0,0). Activate Wi Go to Settings t