The popularity of text-to-image artificial intelligences could be their own downfall: if the pictures they produce proliferate too much, they could contaminate the data sets that new models are trained on, harming performance.
AI tools like DALL-E 2, Midjourney and Stable Diffusion can create pictures of whatever people request, and their images are increasingly being shared online.
Such AIs improve their results by training on data sets of and associated captions from the internet. To investigate how the possibility of AIs consuming their own output may affect training, at the RIKEN Advanced Data Science Project in Japan and his colleagues made almost 2 million images using the Stable Diffusion AI, using industry-standard category names as text inputs to prompt the creation of the images.
Advertisement
The researchers then randomly substituted 20 per cent, 40 per cent and 80 per cent of real images in common data sets with their created images, before training a new AI model on each of these increasingly 鈥渃ontaminated鈥� collections.
AIs trained on the unadulterated data set put never-seen-before images in the right category 鈥� out of a choice of 1000 categories 鈥� 75.6 per cent of the time. This dropped to 74.5 per cent when 20 per cent of the data set was made up of fake images, 72.6 per cent with 40 per cent fake images and 65.3 per cent with 80 per cent fake images.
The team also showed that the AIs produced lower-quality pictures when their training included AI-generated images, by using a metric called the Fr茅chet inception distance to assess how much the output is like real images.
The findings show that it is vital for future AI models to be trained on data sets that are themselves free of AI-generated images, write the researchers.
at the Alan Turning Institute in London says that the level of AI-generated images published online would have to grow enormously before it became a problem, but that safeguards would be a wise precaution.鈥淚 think this illustrates very clearly that all AI systems require human oversight and monitoring,鈥� she says. 鈥淭here鈥檚 always a need for humans to review the outputs of models and to revisit processes of training data selection and data curation.鈥�
Reference:
