sock, feet glove, hand

is an exception - in general might not be a super interpretable in the embedded space

as model gets better, the probability you are falsely recognized becomes more of a threat because we place more trust in these technologies even if there is a small chance they are wrong

government and trust, we lose a lot of our private and personal life because that information is owned by tech companies and government.

f^2 is identity does not mean that the entire layer does nothing. It only means that the activation function does nothing (f(z) = z) so a new linear transformation corresponding to the second layer is still being applied and it could push the output out of the (0,1) range making it invalid for NLL

REMEMBER: A standard layer in a neural network does two distinct jobs

1) Linear Transformation involving weights and bias. - z = w.Tx+b

2) Activation Function then passes the new value (z) through an activation function to get the final output of the layer f(z) = a

This model is ill-formed because if it is a one layer identity network, the network computes w.Tx+b and the network doesn't alter this linear combination at all. It just passes straight through, meaning that this network can output any real number from -infinity to infinity.

The Negative Log-Likelihood takes a probability between 0 and 1. This MISMATCH is why the network breaks down.