Consider a fully connected autoencoder (each hidden node is connected to all inputs and all outputs)...

Transcribed Image Text:

Consider a fully connected autoencoder (each hidden node is connected to all inputs and all outputs) with two dimensional binary input and one hidden layer with softplus activation function. At iteration t, the weights are shown in the following autoencoder architecture along with the input vector (x1=0, x2=1). All bias values are zero. Assume learning rate = 0.25 and momentum constant = 0.75. Also assume at t-1, w1= -0.5, w2-0.5, w3-0.5 and w4= -0.5. w3=1 w4-0 * wl=0 w2=1 1. What activation function will you choose at the output node? 2. What loss function will you choose for training this autoencoder? What will be the value of loss function at iteration t? 3. What will be the weights w4 and w2 in iteration t+1 assuming backpropagation with ordinary gradient descent is used? 4. What will be the weights w4 and w2 in iteration t+1 assuming backpropagation with momentum based gradient descent is used? B. Calculate the KL divergence KLD(p(x,y) || q(x,y)) of two continuous distributions p(x,y) and g(x,y) in bits where p(x, y) = 1 24 m (x-3)² 2+9 q(x, y) = — — — e 2 (y-4)² 2-16 12²-2² Consider a fully connected autoencoder (each hidden node is connected to all inputs and all outputs) with two dimensional binary input and one hidden layer with softplus activation function. At iteration t, the weights are shown in the following autoencoder architecture along with the input vector (x1=0, x2=1). All bias values are zero. Assume learning rate = 0.25 and momentum constant = 0.75. Also assume at t-1, w1= -0.5, w2-0.5, w3-0.5 and w4= -0.5. w3=1 w4-0 * wl=0 w2=1 1. What activation function will you choose at the output node? 2. What loss function will you choose for training this autoencoder? What will be the value of loss function at iteration t? 3. What will be the weights w4 and w2 in iteration t+1 assuming backpropagation with ordinary gradient descent is used? 4. What will be the weights w4 and w2 in iteration t+1 assuming backpropagation with momentum based gradient descent is used? B. Calculate the KL divergence KLD(p(x,y) || q(x,y)) of two continuous distributions p(x,y) and g(x,y) in bits where p(x, y) = 1 24 m (x-3)² 2+9 q(x, y) = — — — e 2 (y-4)² 2-16 12²-2²