Без заголовка

Как запоминаемость, посчитанная дешёвым машинным интеллектом, тождественна (до некоторой степени, конечно -- корелляция 0.68 в случае обучения сетки MemNet на датасете LaMem) запоминаемости обычным мозгом обывателя, расписано в исходной серьёзной статье -- в посте есть ссылка. We conducted memorability experiments using Amazon’s Mechanical Turk (AMT) on the 60, 000 target images obtained by sampling the various datasets mentioned in Sec. 2.1. Each task lasted about 4.5 minutes consisting of a total of 186 images divided into 66 targets, 30 fillers, and 12 vigilance repeats. Targets were repeated after at least 35 images, and at most 150 images. Vigilance repeats were shown within 7 images from the first showing. The vigilance repeats ensured that workers were paying attention leading to a higher quality of results. Workers who failed more than 25% of the vigilance repeats were blocked, and all their results discarded. Further, we used a qualification test to ensure the workers understood the task well. We obtained 80 scores per image on average, resulting in a total of about 5 million data points. Similar to [13], we use rank correlation to measure consistency. Comparison with [13]: Before describing the results on our new dataset, we first compare the performance of our method to the one proposed by Isola et al [13] on the SUN memorability dataset to ensure that our modifications are valid. We randomly selected 500 images from their dataset, and collected 80 scores per image. After applying our algorithm to correct the memorability scores, we obtained a within-dataset human rank correlation of 0.77 (averaged over 25 random splits), as compared to 0.75 using the data provided by [13]. Further, we obtain a rank correlation of 0.76 when comparing the independently obtained scores from the two methods. This shows that our method is well suited for collecting memorability scores. Заодно там считали отношение запоминаемости с популярностью, "фиксацией" (наличие характерных особенностей), эмоциями и эстетикой. Например, We find that images that evoke disgust are statistically more memorable than images showing most other emotions, except for amusement. Further, images portraying emotions like awe and contentment tend to be the least memorable. This is similar to the findings in [12] where they show attributes like ‘peaceful’ are strongly negatively correlated with memorability. Overall, we find that images that evoke negative emotions such as anger and fear tend to be more memorable than those portraying positive ones.