Training new AI models is more accurate when the training data is generated by another AI model as opposed to real data.

Training a generative AI model is best done using data generated by another AI model. Researchers from MIT and Google came to this conclusion. They focused on models that can generate images. Such models are trained using huge datasets containing labeled images. An AI model first learns what a cat looks like by looking at many pictures that contain cats. Traditionally, the training data comes from real life. In other words: researchers work with real photos.

Learning from an AI

The researchers have now discovered that this is not the best method. Today there are already good generative AI models that can generate images. The results of these models appear to be more suitable for training new AI systems. In other words, AIs in school prefer to be taught by other, more experienced AIs.

The researchers trained a model with twenty million AI-generated images and contrasted it with another model that acquired its knowledge based on fifty million real-world images. Even though the synthetic image dataset was less than half the size, the model performed better after training.

More details

The synthetic dataset can contain multiple images that are the result of the same prompt to the existing AI model. This creates multiple correct images that match exactly the same description, which the researchers say allows the learning model to gain more insight.

However, the researchers put their results into perspective somewhat. For example, training with only synthetic images brings its own challenges. Finally, they have lower resolution and are subject to the limitations of the original model. Scientists at Google and MIT expect training with both types of datasets to become the norm in the future.