Language
Country/Area
No result found
From the first generation to the advanced generation: How does early model collapse quietly affect little data?
With the rapid development of machine learning technology, many researchers have begun to notice an emerging phenomenon: model collapse. This phenomenon describes the process by which machine learning models gradually lose information and performance when trained on unorganized synthetic data. According to the definition proposed by Shumailov et al., model collapse is divided into two stages: early model collapse and late model collapse.
In early model collapse, the model begins to lose information about the tails of the data distribution, which mainly affects accuracy on a small number of data.
When we use synthetic data to train models, potential problems surface, including function approximation errors, sampling errors, and learning errors. These problems can occur even on the simplest models, especially in complex models, where errors are more likely to accumulate, accelerating model collapse. This makes early crashes difficult to detect, as overall performance may appear to be improving, while performance for a small number of data is declining.
Late model collapse can result in significant losses in model performance, misconceptions, and loss of much of the variability.
Root cause of model collapse
The root causes of model collapse can be summarized into three aspects, one is function approximation error, the other is sampling error, and the third is learning error. The accumulation of these problems can lead to a decline in the overall performance of the model. Especially in the Internet sharing of data, the data generated by AI will enter the future training data set, causing a vicious cycle.
Many researchers are concerned about this phenomenon and believe that model collapse will pose a fundamental threat to the future development of generative AI. However, some researchers have recently come up with a different perspective. They argue that model collapse is avoidable as long as synthetic data is accumulated alongside human-generated data. Their study points out that the pattern of data accumulation over time is more realistic, rather than just deleting all the data every year.
The real-world impact may not be as pessimistic as one might think.
Exploration of alternative solutions
In addition to the above discussion, another school of literature studies the use of machine learning detectors and watermarking techniques to identify data generated by the model and filter out such data. These methods provide new ideas for dealing with model collapse.
Preliminary exploration of mathematical models
In 2024, researchers made the first attempt to use a simple 1D Gaussian model to demonstrate the collapse phenomenon. The model uses an unbiased estimator based on the original data to calculate the mean and variance. Although the results of this simple model do not fully reflect the complexity of reality, it provides a basis for further research.
As the model evolves, even after the first generation, the complete distribution no longer appears normal, but instead transforms into a variance-gamma distribution.
Although this exploration may seem theoretical, its significance lies in providing a tool to help understand and evaluate changes between different generations. Through these models, researchers can calculate the expected mean and variance in each generation to better understand the dynamics of model collapse.
In the face of the reality that machine learning models are gradually gaining popularity, we should further think about: Will future generative AI be able to successfully cope with the challenge of model collapse, or will it unknowingly fall into a deeper dilemma?
