A machine learning engineer is developing a Generative Adversarial Network (GAN) to produce synthetic, high-resolution images of circuit boards for augmenting a small dataset. During the training process, the engineer observes that the generator network is producing a very limited set of highly similar, but realistic-looking, images. Despite trying different random noise inputs, the output variety does not increase. The discriminator's loss decreases, but the generator's loss stagnates, indicating it has found a few outputs that consistently fool the discriminator.
Which of the following GAN-specific training failures is the engineer most likely experiencing?
The correct answer is mode collapse. This is a common failure mode in GAN training where the generator learns to produce only a limited variety of samples (or modes) from the target distribution, rather than capturing its full diversity. The scenario described, where the generator produces a small set of similar images regardless of the input noise, is a classic example of mode collapse.
The vanishing gradient problem occurs when the discriminator becomes too effective, providing gradients to the generator that are too small to facilitate effective learning. While it is a GAN training issue, it describes the mechanism of learning failure, not the specific symptom of low-variety outputs.
Concept drift refers to a change in the statistical properties of the target variable over time, which is a problem that affects models in production, not a specific internal failure mode during GAN training.
Discriminator overfitting happens when the discriminator memorizes the training set. While this can lead to other problems, including causing mode collapse, the phenomenon of the generator producing a limited variety of outputs is specifically termed mode collapse.
Ask Bash
Bash is our AI bot, trained to help you pass your exam. AI Generated Content may display inaccurate information, always double-check anything important.
What exactly is mode collapse in GANs?
Open an interactive chat with Bash
Why does mode collapse happen during GAN training?
Open an interactive chat with Bash
How can mode collapse be mitigated in GAN training?