A new study warns that AI-generated models could collapse as they increasingly rely on AI-generated content for their training.
Artificial Intelligence (AI) models could face a new problem soon as AI-generated content increasingly spreads across the Internet. The great linguistic models (LLM), like ChatGPT from OpenAI, have relied on data available online to train and improve their models.
However, as these models exhaust data available online or face greater restrictions on data access, they can be trained with AI-generated content.
According to a new study, this could lead to a performance degradation of the models and would ultimately lead to the production of incoherent content, a phenomenon known as ‘model collapse’.
“Over time, we anticipate it will be more difficult to train the modelsalthough we probably have more data, simply because it is very easy to sample model data“Ilia Shumailov, junior researcher at the University of Oxford and co-author of the study, explains to ‘Euronews Next’.
“But what’s going to happen is that it’s going to be harder to find a population of data that is not really biased,” he added. The study, published in the journal ‘Nature’, analyzes what happens when models train with AI-generated data across multiple cycles.
The research found that systems begin to make significant mistakes and fall into meaninglessness. another article by Emily Wenger, researcher at the Duke University, demonstrates this through an experiment in which an AI model is continuously trained with content generated by AI.
In the experiment, an AI model was provided with a data set containing photos of different breeds of dogswith a overrepresentation of golden retrievers.
The study found that the model was more likely to generate images of golden retrievers than other less-represented dog breeds. As the cycle progressed, the model was leaving aside other races until it started generating meaningless images.
Stages of ‘model collapse’ in Artificial Intelligence
“The collapse of the model has two stages. The first is what we call the initial stage of model collapseand what happens is that when one model learns from another, the first thing you see is a reduction in variance,” explains Shumailov.
This results in an oversampling of certain aspects, while other important aspects are simply neglected. because they were not entirely clear for the initial model.
This is when AI models stop being useful because previous models introduce their own errors in the data. The errors present in the initial data are passed to the next model, which adds its own set of errors and transmits it too.
As data is continually produced and recycled, models begin to misinterpret reality and make more mistakes.
“If there are some errors within the data that were generated by model one, they basically propagate to the next model. And ultimately, The result is that the model basically misinterprets reality“, explica Shumailov.
Types of errors in AI models
According to Shumailov, there are three types of errors that models can make: architectural errors, learning process errors and statistical errors.
Architectural errors occur when the structure of the AI model is not adequate to capture all the complexities of the data provided to it, causing inaccuracies because the model misinterprets or oversimplifies some parts.
ADVERTISINGLearning process errors occur when the methods used to train models have inherent biaseswhich pushes the model to make certain types of errors.
Implications of ‘model collapse’
When models collapse, the main concern is that Slow down the pace of your performance improvement. AI models are highly dependent on the quality of the data they are trained on.
However, when trained with AI-generated content, this data continually introduces errors in the system.
“We will probably have to spend additional effort filtering the data. And this will probably involve a slowdown in improvement“, afirma Shumailov.
ADVERTISINGFurthermore, as the variance decreases and the data becomes less diverse, it is expected that underrepresented data are disproportionately affected, what it poses doubts about the inclusivity of AI models.
“We have tor extremely careful to make sure our models are fair and don’t lose sight of the minority data they contain,” says Shumailov.
