Skip to content

13.1-Combined.md

Latest commit

 

History

History
197 lines (117 loc) · 6.1 KB

13.1-Combined.md

File metadata and controls

197 lines (117 loc) · 6.1 KB

Quiz: Generative Models - Deep Dive

Question 1 (Multi-Select)

Which of the following accurately describe the role and goals of generative models in unsupervised learning?

  • A. They only classify input data into known categories.

  • B. They estimate the probability distribution over the input space.

  • C. They can be used to generate new, artificial samples resembling the data.

  • D. They are used for tasks such as density estimation.

  • E. They require labeled data to function effectively.

Show Answer

Correct Answers: Explanation:
Generative models do not require labeled data and do more than classification.

"Here we take a probabilistic view of unsupervised learning and try to estimate the probability distribution over the input space."
"In this lesson, we'll focus on Density Estimation... we may want to just have the ability to generate samples from this distribution..."

Question 2 (Multiple Choice)

Which traditional method was used for density estimation before the resurgence of deep learning, despite its limitations with high-dimensional data?

  • A. Naive Bayes classifiers

  • B. Support Vector Machines

  • C. Gaussian Mixture Models

  • D. Linear Discriminant Analysis

Show Answer

Correct Answers: Explanation:
GMMs were traditionally used, though they struggle with high-dimensional input spaces.

"For example, Gaussian mixture models also produce some estimate of the probability distribution over the input space. However, these methods have severe deficiencies when the input is very high dimensional."

Question 3 (True/False)

Generative models model the conditional probability of labels given the input data.

  • A. True

  • B. False

Show Answer

Correct Answers: Explanation:
Discriminative models model ( P(y|x) ); generative models model ( P(x) ).

"Discriminative models, model the conditional distribution probability of the label given the input... Generative models, on the other hand, model the distribution over the input space."

Question 4 (Multi-Select)

What are key challenges associated with modeling the joint distribution of high-dimensional data in generative models?

  • A. High computational complexity

  • B. Intractability of the exact probability distribution

  • C. Lack of any theoretical basis

  • D. Requirement of assumptions or simplifications

Show Answer

Correct Answers: Explanation:
Modeling ( P(x) ) directly is challenging without simplifying assumptions.

"This is a very intractable and hard thing to do. And so we'll have to make various assumptions or simplifications in order to make this feasible."

Question 5 (Concept Application)

You are asked to design a generative model for a new high-dimensional image dataset. What strategies from the transcript are likely to help in tackling the complexity?

  • A. Use deep learning architectures that reduce dimensionality

  • B. Use rule-based symbolic reasoning

  • C. Leverage embeddings to find low-dimensional representations

  • D. Avoid simplifications to maintain full generality

Show Answer

Correct Answers: Explanation:
The transcript emphasizes dimensionality reduction using neural networks.

"Deep learning is very good at learning features that extract meaningful information in a low dimensional embedding from high dimensional data."

Question 6 (Multiple Choice)

Which of the following statements best captures the distinction between discriminative and generative models?

  • A. Discriminative models generate new samples from the data distribution.

  • B. Generative models are always more accurate than discriminative ones.

  • C. Discriminative models estimate ( P(y|x) ), while generative models estimate ( P(x) ).

  • D. Generative models require supervised data, while discriminative models do not.

Show Answer

Correct Answers: Explanation:
This is a central distinction made in the lesson.

"Discriminative models, model the conditional distribution probability of the label given the input... Generative models... model the distribution over the input space."

Question 7 (Multi-Select)

According to the transcript, why is deep learning suitable for generative modeling of high-dimensional data?

  • A. Deep learning methods can extract meaningful low-dimensional features.

  • B. Neural networks can model complex distributions with sufficient flexibility.

  • C. They inherently encode explicit probabilistic rules.

  • D. They eliminate the need for any data preprocessing.

Show Answer

Correct Answers: Explanation:
Deep learning reduces dimensionality and provides flexibility in modeling.

"Deep learning is very good at learning features that extract meaningful information in a low dimensional embedding..."
"Just, like discriminative models we can have a parametric approximation of this distribution."

Question 8 (True/False)

The ability to generate new data samples is a core feature of generative models.

  • A. True

  • B. False

Show Answer

Correct Answers: Explanation:
Generating samples is one of the core applications.

"We may want to just have the ability to generate samples from this distribution, that is actually generate artificial examples..."

Question 9 (Multiple Choice)

Which principle is often used to optimize the parameters of generative models?

  • A. Cross-entropy minimization

  • B. Maximum likelihood estimation

  • C. Mean squared error

  • D. KL divergence minimization

Show Answer

Correct Answers: Explanation:
MLE is the optimization principle used for learning generative models.

"We can have a set of parameterized models ( p(x, \theta) )... use the principle of maximum likelihood