How deeply does How Large Language Models Work cover mode collapse?

Explains the mode, distribution bias, and repeated-training mechanism, but does not provide prevention methods.

Mode collapse

As explained in How Large Language Models Work

Mode collapse is a generative-model failure in which outputs become concentrated on a small set of common patterns while rare or diverse patterns are lost.

Mode collapse is a generative-model failure in which outputs increasingly favor the most common patterns in the training data while losing rare, nuanced, and interesting ones. A model may produce fluent results, yet represent less of the original variety. section 9.4.2

The failure develops through a feedback loop:

A generative model emphasizes the mode—the most common value or pattern in its training distribution. section 9.4.2
Rare or nuanced cases appear less often in generated output than they did in the original data. section 9.4.2
Training another model on that output reinforces the common patterns and removes more of the distribution’s variety. section 9.4.2
After enough repetitions, the model can become uselessly repetitive, producing nearly the same kind of output. section 9.4.2

A shrinking recipe collection

Imagine a recipe collection containing familiar dishes and a few unusual regional recipes. A model trained on it favors the familiar dishes. If its generated recipes become the next model’s training data, the unusual recipes become still less visible until the collection effectively contains only common dishes. The same pattern can affect text or images when repeated generations replace more varied source data. section 9.4.2

How How Large Language Models Work covers it

liveBook manning.com

Read How Large Language Models Work on liveBook Get the book at manning.com

Also explained in

liveBook manning.com

Inside Deep Learning: Math, Algorithms, ModelsA GAN failure mode in which the generator produces only a limited variety of outputs instead of covering the data distribution.
liveBook manning.com

Generative AI in ActionMode collapse is a generative-model failure in which the model produces a narrow range of outputs repeatedly instead of representing the diversity of the data distribution.
liveBook manning.com

Evolutionary Deep Learning: Genetic algorithms and neural networksMode collapse is a generative-adversarial-network failure in which the generator produces only one or a small range of output variations despite the diversity of the training data.

Compare how 4 books explain mode collapse →

Related concepts

noise additionaddressed by · Noise addition is the deliberate introduction of random perturbations into data, model representations, outputs, or objectives.
Wasserstein distanceaddressed by · Wasserstein distance is a distance between probability distributions defined by the minimum cost of transporting probability mass between them under a specified ground distance.
generative adversarial networkuses · A generative adversarial network is a generative model with a generator and discriminator trained in competition, enabling the generator to produce samples resembling a target dataset or distribution.
WGAN-GPaddresses · WGAN-GP is a Wasserstein generative adversarial network variant that adds a gradient penalty to constrain the critic and stabilize training.
Wasserstein GANaddressed by · A Wasserstein GAN is a generative adversarial network that uses a critic and a Wasserstein-based objective to compare generated and real-data distributions.

This explainer was generated by Manning's knowledge graph from the text of How Large Language Models Work. The section links point to the parts of the book it draws on.