Model Collapse
The situation that models lose diversity and quality as a result of retraining with data made by artificial intelligence.
As a result of the training of model crash (model crash), artificial intelligence models, real human production data is produced by artificial intelligence, gradually increasing in place of data, it defines the situation of losing its ability to represent diversity, quality and real world distribution over time. This process consists of the training data of the next model of the output of a model, and the repeating of this loop: the model in each repetition is rare in the tail of the data distribution, but the actual examples are somewhat forgotten and the most likely, the "average" is close to the nucleus; the content produced in the result is increasingly uniform, the general pass, and it becomes lack of interesting details.
This phenomenon has become an important concern source because the internet is increasingly transmitted with the content of artificial intelligence production, because the educational data of future models may contain the outputs of previous models at an increased rate without regard, which leads to a problem called "data pollution". To prevent model crash, researchers focus on protecting the share of actual, verified human production content in education data, labeling and monitoring synthetic and real data, and developing systems that follow the origin of data source. This topic becomes a increasingly critical research field in terms of sustainability of model training with the prevalence of manufacturer artificial intelligence.
This phenomenon has become an important concern source because the internet is increasingly transmitted with the content of artificial intelligence production, because the educational data of future models may contain the outputs of previous models at an increased rate without regard, which leads to a problem called "data pollution". To prevent model crash, researchers focus on protecting the share of actual, verified human production content in education data, labeling and monitoring synthetic and real data, and developing systems that follow the origin of data source. This topic becomes a increasingly critical research field in terms of sustainability of model training with the prevalence of manufacturer artificial intelligence.