Teach AI models to learn the same way as infants learn

Abstract

Understanding how infants develop their visual perception provides valuable insights into how artificial intelligence models can be trained to learn visual representations more effectively. This study explores the impact of training self-supervised AI models using biologically inspired methods that mimic infant visual development. Specifically, training data was progressively altered to simulate an infant’s developing visual acuity and colour perception. The CO3D dataset, which provides multi-view images of objects, was used to approximate how infants view their environment through interaction and exploration. Two self-supervised learning models, DINO and MAE, were trained using blurred and desaturated images, gradually transitioning to clear and fully coloured images over time. The trained models were evaluated on object classification (CO3D and ImageNet) and image segmentation (COCO). Results suggest that gradual exposure to varied visual conditions can influence AI’s ability to learn visual representations, depending on the learning algorithm used and the respective task. These findings indicate that integrating biological learning principles into AI model training could enhance performance.