Abstract
Image memorability can be defined as a combination of intrinsic characteristics of the image itself and the degree of its correspondence to the human observer’s knowledge structure of the world. Understanding and extracting the image features that contribute to memorability is still a challenge. Advances in computational techniques have allowed for interpreting what makes an image memorable, in addition to predicting the memorability of a given image. Visual memory schemas (VMS) are one such operationalization that defines image memorability as two-dimensional memorability maps that capture the most memorable regions of the scene, predicting with a high degree of consistency human observer’s memory for the same images. These maps correlate with mental schemas employed by humans to encode visual memories. Here we ask whether it is possible not only to predict but also modulate human memory with artificially generated images. We developed a computational approach based on deep learning models for estimating and enforcing the VMS maps when generating realistic high-resolution images. The generated images are high and low memorability pairs of images, where the only difference between images is a variation in the continuum of VMS-defined memorability. We then conducted a recognition memory experiment, where human observers are shown sequences of artificially generated images and are asked to indicate if they have seen a given image before. The observers show a significantly superior memory for the highly memorable images compared to poorly memorable images, for VMS-defined memorability and hit rate. Raising memorability of an image also increased the chance of it being falsely remembered, mirroring findings from visual memory schema experiments that employ real images. Implementing and testing a construct from cognitive science allows us to generate realistic images whose memorability we can manipulate at will as well as providing a tool for further study of mental schemas in humans.