No abstract

No abstract

You may find some images easier to remember than others. Recent studies of visual memory have found remarkable levels of consistency for this inter-item variability across observers, suggesting that memorability can be considered an intrinsic image property. The current study replicated and extended previous results, while adopting a more traditional visual long-term memory task with retention intervals of 20 min, one day, and one week, as opposed to the previously used repeat-detection task, which typically relied on short retention intervals (5 min). Our memorability rank scores show levels of consistency across observers in line with those reported in previous research. They correlate strongly with previous quantifications and appear stable over time. Furthermore, we show that the way consistency of memorability scores increases with the number of responses per image follows the Spearman-Brown formula. Interestingly, our results also seem to show an increase in consistency with an increase in retention interval. Supported by simulated data, this effect is attributed to a decrease of extraneous influences on recognition over time. Finally, we also provide evidence for a log-linear, rather than linear, decline of the raw memorability scores over time, with more memorable images declining less strongly.

Images differ in their memorability in consistent ways across observers. What makes an image memorable is not fully understood to date. Most of the current insight is in terms of high-level semantic aspects, related to the content. However, research still shows consistent differences within semantic categories, suggesting a role for factors at other levels of processing in the visual hierarchy. To aid investigations into this role as well as contributions to the understanding of image memorability more generally, we present MemCat. MemCat is a category-based image set, consisting of 10K images representing five broader, memorability-relevant categories (animal, food, landscape, sports, and vehicle) and further divided into subcategories (e.g., bear). They were sampled from existing source image sets that offer bounding box annotations or more detailed segmentation masks. We collected memorability scores for all 10 K images, each score based on the responses of on average 99 participants in a repeat-detection memory task. Replicating previous research, the collected memorability scores show high levels of consistency across observers. Currently, MemCat is the second largest memorability image set and the largest offering a category-based structure. MemCat can be used to study the factors underlying the variability in image memorability, including the variability within semantic categories. In addition, it offers a new benchmark dataset for the automatic prediction of memorability scores (e.g., with convolutional neural networks). Finally, MemCat allows the study of neural and behavioral correlates of memorability while controlling for semantic category.

We introduce a framework that uses Generative Adversarial Networks (GANs) to study cognitive properties like memorability, aesthetics, and emotional valence. These attributes are of interest because we do not have a concrete visual definition of what they entail. What does it look like for a dog to be more or less memorable? GANs allow us to generate a manifold of natural-looking images with fine-grained differences in their visual attributes. By navigating this manifold in directions that increase memorability, we can visualize what it looks like for a particular generated image to become more or less memorable. The resulting "visual definitions" surface image properties (like "object size") that may underlie memorability. Through behavioral experiments, we verify that our method indeed discovers image manipulations that causally affect human memory performance. We further demonstrate that the same framework can be used to analyze image aesthetics and emotional valence. Visit the GANalyze website at http://ganalyze.csail.mit.edu/.