Week 6: Deep Generative Neural Networks

Question 1

Generative Adversarial Network (GAN)

Answer 1

A machine learning model for creating outputs that resemble the inputs (i.e. image generation). The two mechanisms of GAN’s are generators, which attempts to build an example to fool the Discriminator, which tries to discern whether a given example is from the input data or an original example made by the generator. The iterations are nested such that the outer loop updates the Generator, and the inner loop updates the Discriminator. This helps prevent overfitting, and the overall objective of creating convincing examples that are difficult to distinguish from the input data.

Question 2

Maximum Likelihood Estimation

Answer 2

When you try to find the probability that best approximates the distribution of datapoints of the input.

Question 3

Density and Sample Estimation

Answer 3

The purpose of this task is to generate samples that best approximates the distribution of training examples.

Question 4

Trackable Density

Answer 4

Generating pixels starting from one corner and working your way towards another corner. The computation is chain and probability-based and it doesn’t scale well.

Question 5

Autoencoder

Answer 5

A type of artificial neural network that encodes the input into a more efficient latent space and then decodes it into the output. Each sample in the input is mapped to one point in the latent space.

Question 6

Variable Autoencoder (VAE)

Answer 6

A type of autoencoder that maps the input into a distribution in the latent space, instead of a single point.