Search Papers | Poster Sessions | All Posters
Poster A68 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Modeling Visual Memorability Assessment: A Computational Approach Using Autoencoders
Elham Bagheri1,2, Yalda Mohsenzadeh1,2 (); 1Vector Institute for Artificial Intelligence, Toronto, ON M5G 0C6, Canada, 2Department of Computer Science, Western University, London, ON N6A 3K7, Canada
Certain images stick in our mind while others vanish quickly. This study explores the computational aspects of image memorability by employing a pretrained autoencoder, specifically a VGG-based model trained on the ImageNet dataset. The research investigates the relationship between the memorability of images, quantified as the likelihood of their remembrance after a single exposure, and their reconstruction errors and distinctiveness in the latent space of an autoencoder, finetuned on the MemCat dataset comprising 10,000 images in diverse categories. The predictive power of latent representations in determining memorability is also evaluated. The findings suggest that images with unique features that challenge the autoencoder’s capacity are inherently more memorable. This correlation indicates a new pathway for evaluating image memorability, potentially impacting industries reliant on visual content and fostering advancements in the fields of artificial intelligence and cognitive science. It also demonstrates how machine learning models can emulate human cognitive processes to assess memorability, leading to improvements in algorithmic performance.
Keywords: image memorability latent code representation autoencoders reconstruction error