Teaching AI to see the world more like we do

Google DeepMind · 8 months ago

Researchers developed a method to reorganize how AI vision models represent visual information so they align better with human perception. The technique uses a three-step process involving a teacher model trained on the THINGS dataset of human judgments, which generates millions of synthetic training examples to fine-tune student models. Models trained this way show improved performance on cognitive tasks, few-shot learning, and robustness to image distribution shifts.