mode collapse

Generative adversarial networks

A failure of generative adversarial networkgenerative adversarial network, n.: Two neural networks trained as rivals: one forges fake images, the other tries to spot the fakes. As both improve, the forger's output becomes realistic enough to fool people. training in which the generator's output loses variety: large regions of its noise input all yield the same or nearly the same sample, so only a few modes (regions of high probability) of the data distribution are ever produced. It arises because the generator is rewarded for fooling the current discriminator, not for diversity: it can chase whichever outputs the discriminator currently misjudges, and the two networks may cycle between modes instead of converging. Partial collapse — reduced variety rather than a single output — is common and is also called mode dropping. Remedies include minibatch discrimination, unrolled objectives, Wasserstein lossesWasserstein GAN, n.: A version of the forger-versus-detective image generator that measures how far its fakes are from real data in a smoother way, making training steadier and easier to monitor., spectral normalization and larger batches; diffusion modelsdiffusion model, n.: An AI that learns to create images, sound or video by practicing noise removal: shown pictures buried under static, it learns to clean them up, then builds new ones starting from pure static. and likelihood-trained models are far less prone to it.

After epoch 40 every sample was the same brown-haired face: classic mode collapse.

Language models

By extension, the loss of output diversity in a large language modellarge language model, n.: A computer program trained on enormous amounts of text to predict the words that come next. At vast scale, that simple skill lets it chat, answer questions, translate, summarize and write code. after instruction tuninginstruction tuning, n.: Training a language model on many tasks written as plain instructions with correct answers, so it learns to do what a request asks instead of just continuing the text. or reinforcement learning from human feedbackreinforcement learning from human feedback, n.: A way of training AI using people's judgments: humans compare pairs of answers, a scoring model learns their preferences, and the AI is then trained to produce answers that score well.: the model concentrates probability on a few favored answers, phrasings or styles — naming the same 'random' number, writing near-identical stories — even when variety is wanted. It differs from a low sampling temperaturetemperature, n.: A setting that controls how adventurous an AI model's word choices are. Low temperature makes it pick the safest, most likely words; high temperature makes it more varied and surprising, but more error-prone. because the preference persists across prompts and sampling settings.