AL3451_Machine Learning Assignement 2

Which learning paradigm best describes kNearest Neighbors (kNN)?

• Eager learning

• Lazy learning

• Reinforcement learning

• Online learning

In kNN, the value of k primarily controls …

• Model bias

• Model variance

• Learning rate

• Feature scaling

A drawback of basic kNN is that it becomes inefficient when …

• The number of classes is small

• The training set is large and highdimensional

• Continuous features are zscore normalized

• Euclidean distance is used

Which distance metric is most robust to outliers in kNN?

• Manhattan (L1)

• Euclidean (L2)

• Minkowski with p = 3

• Cosine

Gaussian Mixture Models (GMMs) assume that the data distribution is a …

• Single multivariate Gaussian

• Deterministic manifold

• Weighted sum of Gaussian components

• Product of independent Gaussians

In a GMM, the mixing coefficients must satisfy which property?

• All positive, sum to 1

• All nonnegative, sum to 1

• Any real value, sum to 0

• Any real value, unrestricted sum

ExpectationMaximization (EM) alternates between steps that respectively …

• Optimize latent variables, then hyperparameters

• Minimize loss, then maximize entropy

• Estimate posterior membership, then reestimate parameters

• Shuffle data, then update gradients

Which of the following will always increase (or leave unchanged) after each EM iteration on a GMM?

• KullbackLeibler divergence

• Loglikelihood

• Number of clusters

• Model entropy

The soft assignments computed in the Estep of EM for GMMs are known as:

• Prior probabilities

• Responsibilities

• Hyperparameters

• Likelihoods

When fitting a GMM, choosing too many components mainly risks:

• Underfitting

• Mode collapse

• Overfitting

• Vanishing gradients