The k-means algorithm for clustering is guaranteed to converge to a local optimum. It's possible for k-means and Gaussian mixture model algorithms to converge to different cluster centers but still give the same clustering assignment. (Hint: exam practice question) PCA is an unsupervised approach, as such it is possible that the projected data might be more difficult to classify. Training with dropout is similar to training multiple neural networks and averaging their outputs during test time. When a decision tree is grown to full depth, it is leas likely to fit the noise in the data. Q2) Multiple choice ( 16pts ) - Asguming that you're using gigmoid activation function, this kind of weight initialization will cauge all the neurons to get saturated at either -1 or 1. Gradients will be all zero. What kind of initialization could this be? Choose from the options below: a) initialization with very large numbers b) initialization with very small numbers c) initialization with all zeros - In regression, the labels are a) unknown b) members of a discrete set c) real numbers d) any of the above - As the regularization parameter for linear regresgion increases, the training error must: a) increase b) decrease c) neither A nor B are guaranteed - As the regularization parameter for linear regresgion increases, the teating error must: a) increase b) decrease c) neither A nor B are guaranteed - Which of the following atatements about Naive Bayes is incorrect? a) Attributes are equally important. b) Attributes are atatiatically dependent of one another given the class value. c) Attributes are statistically independent of one another given the class value. d) Attributes can be nominal or numeric e) All of the above