Machine Learning Basics
Which of the following correctly describes models that generate images and similar data?
- AA GAN (generative adversarial network) trains only the discriminator, which tells real data from fakes, and does not train the generator, which produces the data
- BA diffusion model produces an image by performing the noise-removal computation only once, so it needs less computation to generate than a GAN
- CA diffusion model considers a process that adds noise to data little by little, learns to remove that noise step by step, and generates images from noise
Show answer and explanation
Answer: C. A diffusion model considers a process that adds noise to data little by little, learns to remove that noise step by step, and generates images from noise
A GAN trains two networks alternately so that they compete with each other: a generator that produces data and a discriminator that judges whether data is real or made by the generator. A diffusion model learns a network that estimates and removes the noise at each step, so as to reverse the process of adding noise to data little by little, and generates data step by step from pure noise. A diffusion model needs computation over many steps to generate, but its training tends to be stable and it can produce diverse, high-quality images.
- A. A GAN trains both the generator and the discriminator, alternately.
- B. A diffusion model repeats the noise-removal computation over many steps, so generation generally needs more computation than with a GAN.
- C. Correct. It learns to remove noise step by step and produces data from noise.