MIT Researchers Show Synthetic Data Can Rival Real Images for AI Training
MIT researchers develop a generative model that creates synthetic data to train AI for image classification, potentially saving millions in dataset costs. The method sometimes outperforms models trained on real data.
MIT researchers discover that artificial intelligence can learn to identify images using synthetic data instead of real-world examples. The team creates a special generative machine learning model that produces highly realistic synthetic images, which then successfully train another AI model for vision-related tasks. This approach challenges the traditional requirement of using massive, expensive datasets of real images to teach machines how to classify objects.
Currently, training an AI to perform tasks like identifying damage in satellite photos requires enormous datasets that cost millions of dollars to compile. The MIT team's generative model requires significantly less memory to store and share compared to these massive datasets. When the researchers compared their synthetically trained model to other models trained on real data, they find that their method learns visual representations equally well and sometimes even performs better than the real thing.
The training process works by showing the generative model millions of images of a specific object class so it learns to generate similar objects in various scenarios. For example, after learning from pictures of cars, the model imagines and creates images of cars in different poses, colors, and sizes that it never observed during training. The researchers then connect this pretrained generative model to a contrastive learning model, which uses these multiple varied views to effectively learn which image pairs are similar or different.