AI models collapse when trained on recursively generated data.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Generative Models

Background:

• Generative artificial intelligence (AI), including large language models (LLMs) like GPT-4 and image generation models like stable diffusion, is rapidly transforming online content.
• The widespread use of AI-generated text and images raises questions about the future of data used for training these models.
• Previous AI advancements, such as GPT-2, GPT-3(.5), and GPT-4, have shown significant capabilities in various language tasks.

Purpose of the Study:

• To investigate the potential impact of large language models (LLMs) on future AI training data.
• To identify and analyze the defects that arise when AI models are trained on model-generated content.
• To understand the implications of these defects for the sustainability of AI development and the value of diverse data sources.

Main Methods:

• Theoretical analysis of generative models, including LLMs, variational autoencoders (VAEs), and Gaussian mixture models (GMMs).
• Simulation and empirical studies to demonstrate the occurrence and effects of model collapse.
• Investigation into the loss of data distribution tails when models are trained on synthetic data.

Main Results:

• Indiscriminate use of model-generated content in training leads to irreversible defects in AI models.
• The phenomenon, termed 'model collapse,' causes the disappearance of the tails of the original data distribution.
• Model collapse is shown to be a ubiquitous issue across various types of generative models, including LLMs, VAEs, and GMMs.

Conclusions:

• Model collapse poses a significant threat to the long-term quality and diversity of AI-generated content.
• Sustaining the benefits of training on large-scale web data requires addressing model collapse.
• Data reflecting genuine human interactions will become increasingly valuable as a countermeasure to model collapse in AI training.