Chapter 8

1. No, they don't. Both the encoder and decoder must be functionally symmetric, but their internal structures can also be different.
2. No; a part of the input information is lost during the transformation, while the remaining one is split between the code output Y and the autoencoder variables, which, along with the underlying model, encode all of the transformations.
3. As min(sum(z_i)) = 0 and min(sum(z_i)) = 128, a sum equal to 36 can imply both sparseness (if the standard deviation is large) and a uniform distribution with small values (when the standard deviation is close to zero).
4. As sum(z_i) = 36, a std(z_i) = 0.03 implies that the majority of values are centered around 0.28 (0.25 ÷ 0.31), the code can be considered dense.
5. No; a Sanger network (as well as a Rubner-Tavan one) requires the input samples x_i ∈ X.
6. The components are extracted in descending order...