MS-DS Master of Data science Machine Learning 4 — Questions and Answers
Question 1: What is the purpose of the learning rate in gradient descent optimization?
- It controls how many epochs the model trains for
- It scales the step size taken in the direction of the negative gradient (Correct answer)
- It determines the fraction of data used per mini-batch
- It regularizes the model by penalizing large weights
Correct answer: It scales the step size taken in the direction of the negative gradient
The learning rate is a scalar that multiplies the gradient to determine how large a step is taken toward the loss minimum during each parameter update.
Question 2: Which concept describes the phenomenon where a feature's importance appears artificially low because its effect is shared with correlated features in a model?
- Confounding
- Multicollinearity (Correct answer)
- Heteroscedasticity
- Autocorrelation
Correct answer: Multicollinearity
Multicollinearity among predictors causes their individual coefficients or feature importances to be unstable and potentially misleading because the model cannot isolate their separate contributions.
Question 3: In reinforcement learning, what is the Bellman equation used for?
- Computing the gradient of a policy with respect to reward
- Expressing the value of a state as the immediate reward plus discounted value of successor states (Correct answer)
- Defining the exploration-exploitation tradeoff via epsilon decay
- Approximating the Q-function using a neural network
Correct answer: Expressing the value of a state as the immediate reward plus discounted value of successor states
The Bellman equation decomposes the value function recursively: V(s) = E[r + γV(s')], forming the foundation of dynamic programming and Q-learning.
Question 4: What distinguishes a generative model from a discriminative model?
- Generative models only work with continuous data; discriminative models work with categorical data
- Generative models learn the joint distribution P(X,Y); discriminative models learn P(Y|X) directly (Correct answer)
- Generative models require labeled data; discriminative models do not
- Generative models use gradient descent; discriminative models use closed-form solutions
Correct answer: Generative models learn the joint distribution P(X,Y); discriminative models learn P(Y|X) directly
Generative models model how data is generated (joint distribution), enabling them to synthesize new samples, while discriminative models focus solely on the decision boundary P(Y|X).
Question 5: In XGBoost, what is the role of the 'lambda' (L2) regularization parameter?
- It controls the learning rate shrinkage applied to each tree
- It penalizes the squared magnitude of leaf weights to prevent overfitting (Correct answer)
- It sets the minimum number of samples required to split a node
- It controls the subsampling ratio of training instances per tree
Correct answer: It penalizes the squared magnitude of leaf weights to prevent overfitting
The lambda parameter in XGBoost adds an L2 penalty on leaf weight values in the objective function, shrinking them toward zero to reduce model complexity.
Question 6: Which technique is used in neural architecture search and meta-learning to enable rapid adaptation to new tasks with few examples?
- Transfer learning with frozen layers
- Model-Agnostic Meta-Learning (MAML) (Correct answer)
- Knowledge distillation
- Federated learning
Correct answer: Model-Agnostic Meta-Learning (MAML)
MAML learns an initialization of model parameters that can be quickly fine-tuned to new tasks with just a few gradient steps and few labeled examples.
Question 7: What is the curse of dimensionality's primary implication for distance-based machine learning algorithms like k-NN?
- Distances between points become more meaningful as dimensionality increases
- In high dimensions, all pairwise distances converge, making nearest-neighbor search unreliable (Correct answer)
- Feature normalization becomes unnecessary in high-dimensional spaces
- More dimensions always improve model accuracy by providing richer representations
Correct answer: In high dimensions, all pairwise distances converge, making nearest-neighbor search unreliable
As dimensionality grows, the ratio of maximum to minimum pairwise distances approaches 1, so 'nearest' neighbors are nearly as far away as 'farthest' ones, undermining distance-based reasoning.
MS-DS Master of Data science Machine Learning 4
What is the purpose of the learning rate in gradient descent optimization?