Consider the following sentence from the original GAN paper titled Generative Adversarial Nets

in particular, $G$ must not be trained too much without updating $D$, in order to avoid "the Helvetica scenario" in which $G$ collapses too many values of $\mathbf{z}$ to the same value of $\mathbf{x}$ to have enough diversity to model $p_{\text {data }}$

The sentence uses the name of a scenario called Helvetica. Is it related to AI or chemistry? If AI, then where can I read more about the scenario?


1 Answer 1


The "Helvetica scenario" referenced in the paper is an AI issue related to generative systems failing.

In GANs, this might be due to a generator becoming initially strong against the discriminator by focusing on its best output results so far that cover just a subset of the full dataset - i.e. gradients move the generator towards output of a single best class, suppressing output of classes that the discriminator can spot easily, and encouraging output of classes where the discriminator has more trouble. By the time the discriminator catches up with the generator, it may be too late for the combined training system to learn its way out of this state.

In GANs, the term Helvetica scenario from the paper is more commonly referred to as mode collapse in later literature. If you search for "GAN mode collapse" you will find a lot more information about it. There are a few different improvements to GAN design - loss functions and other hyperparameters - that work to reduce the incidence of mode collapse.

From this answer on Data Science SE, it seems like the authors named the effect after a science parody sketch in the show Look Around You, where the structure of a molecule collapses. The "science" in the sketch is complete nonsense.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .