Difussion Models

Question 1

What is the primary goal of a diffusion model in machine learning?

Answer 1

To generate data by learning to reverse a noising process

Question 2

What describes the forward diffusion process in diffusion models?

Answer 2

It is the process of iteratively adding Gaussian noise to the data until it resembles pure noise.

Question 3

During the backward diffusion (reverse process) in diffusion models, what is the primary task of the neural network model?

Answer 3

To sequentially remove noise from the data to reconstruct the original data.

Question 4

In the context of training a diffusion model using CIFAR-10 dataset, which statement is true regarding the preprocessing of images?

Answer 4

Each image is normalized to have values between -1 and 1.

Question 5

What is the significance of ‘timesteps’ in the training process of a diffusion model?

Answer 5

They represent the discrete steps in which noise is added or removed.

Question 6

What does the U-Net architecture in a diffusion model generally consist of?

Answer 6

An encoder-decoder structure with skip connections between mirrored layers.

Question 7

Which of the following best describes the output from an untrained diffusion model when generating images?

Answer 7

The output images are random noise without any discernible features.

Question 8

What is the purpose of the train_one function in the context of diffusion model training?

Answer 8

To perform a single update on the model using a pair of noisy and less noisy images.