Consider the following. A trivial neural network that takes a single, scalar x as input and a scalar y as output. It has a first layer with weight w₁=1.5, bias b₁=1, and nonlinearity (x) = x^2; and it has a second layer with weight w₂=4, bias b₂=450 and no nonlinearity. You can assume that all weights and biases are scalar, so the output y is a scalar as well. The loss function is L(y', y) = (y'-y)^2. One compute graph for this neural network and loss is the following:

h = x*w₁ + b₁

o = h^2o

y' = o * w₂ + b₂

L = (y'-y)^2

Now use back-propagation to compute, the gradient of the loss for the training sample (x₁=2, y₁=520) relative to w₁ is

a. -192

b. -96

c. -1314

d. -768

The gradient of the loss for the same training sample relative to b₁ is

a. -12

b. -6

c. -288

d. -384