Question 6: Low-dimensional features learned by a multiple hidden-layer (two or more layers) autoencoder with non-linearity have more information than the features learned by a single hidden layer autoencoder without non-linearity. True/False?

Question 7: Projecting the data onto a low-dimensional space in the direction of maximum variance will also ensure class separability in the lower dimensional space. True/False?

Question 8: For the data set belonging to one of two classes (a binary classification data) Fisher Discriminant Analysis (FDA) finds the low-dimensional projection of the high-dimensional data set by maximizing the distance between the two classes in the low-dimensional projection. True/False?

Question 9: Multi-Dimensional Scaling (MDS) only works when the input data }D={x1,x2,,xN} lies in the Euclidean space (i.e.,xi Rd for all i) as it enables us to compute the distance between any pairs of points. True/False?

Question 10: For a data set D={x1,x2,,xN} assume that you have computed the Stochastic Neighborhood Embeddings (SNEs). Now given a new data point x0 how will you compute its SNE projection in the lower dimensional space? Explain in no more than 4 lines.