Teaching AI to see the world like humans

Google DeepMind published research in Nature proposing a method to make the internal representations of vision models closer to human knowledge hierarchies, and claims that aligned models are more robust and generalize better. The research uses the cognitive science odd-one-out task to compare human and model judgments, finding that models often rely on surface features such as background color and texture, leading to choices different from humans. The method has three steps: on the pre-trained vision model SigLIP-SO400M, freeze the backbone, train a small adapter on the THINGS dataset as a teacher, generate the AligNet dataset containing millions of images and millions of human-like judgments, then fine-tune the student model accordingly. Aligned models more often agree with human judgments on cognitive science tasks such as odd-one-out and multi-arrangement, and also perform better on few-shot learning and distribution shift.