EX1: Consider a fully connected neural network in the Figure below: X ->> z[1]=Wx+b[1] a[1] =...

  

Transcribed Image Text:

EX1: Consider a fully connected neural network in the Figure below: X ->> z[1]=Wx+b[1] a[1] = o(z) z[2]=w[2]a[1]+b[2] a[2] = (z) c(y, a[2]) a) Assuming the activation function is a Sigmoid function, write the analytical expressions for derivatives with respect to the weights W, biases b, and input x. b) Assuming the activation function is an Identity function, f(x) = x, what would be the derivatives with respect to all the weights W and biases b? Comment on why this activation function is such a bad choice for neural network learning. EX1: Consider a fully connected neural network in the Figure below: X ->> z[1]=Wx+b[1] a[1] = o(z) z[2]=w[2]a[1]+b[2] a[2] = (z) c(y, a[2]) a) Assuming the activation function is a Sigmoid function, write the analytical expressions for derivatives with respect to the weights W, biases b, and input x. b) Assuming the activation function is an Identity function, f(x) = x, what would be the derivatives with respect to all the weights W and biases b? Comment on why this activation function is such a bad choice for neural network learning.