MIT Launches Artificial Intelligence Simulating Human Brain Memory Process

2019-11-04

The reason why world-famous paintings are considered as such is because they are different from ordinary images. From "Mona Lisa" to "Girl with a Pearl Earring" to "Sunflowers," they can stay in people's minds for a long time. However, artists have always been unable to provide a detailed explanation for why this is the case, apart from giving vague and mysterious answers. Therefore, a new study conducted by the Massachusetts Institute of Technology (MIT) used machine learning to generate images that demonstrate what makes certain portraits or scenes stand out among others. In this study, the images that were most memorable to human subjects generally had bright colors, simple backgrounds, and a prominent central position in the composition. Phillip Isola, an assistant professor of electrical engineering and computer science at MIT, said, "There have been many articles about memorability, but this new approach is different. It allows us to imagine what memorability is and provides us with a visual definition that is difficult to express in words." This work builds upon the earlier model called MemNet, which rates the memorability of images and emphasizes the influence of features in the images on decision-making. MemNet's predictions are based on the results of an online study that ranked 60,000 images according to how well they were remembered by human subjects. The current study utilizes a model called GANalyze, which employs a machine learning technique called generative adversarial networks (GANs). GANalyze consists of three modules: one based on MemNet that opens the "memory knob" of the target image and calculates how to achieve the desired effect, another that performs the transformation of instructions, and a final generator that outputs the resulting image. Researchers configured GANalyze to generate images with different aesthetic and emotional appeals. In the process, they found that images with higher scores in terms of aesthetics and emotions tend to have brighter overall tones, richer composition colors, shallow depth of field, and blurred backgrounds, resembling the most unforgettable photographs. Furthermore, GANalyze can also be used to create memorable graphics that help readers remember information they want to retain, potentially bringing revolutionary changes to education. Currently, GANalyze has begun to be used to generate synthetic real-world images to assist in training automated systems to recognize places or objects that are unlikely to be encountered in real life. Alex Andonian, one of the authors of the study, stated that generative models provide new and creative ways for collaboration between humans and machines, which is why he chose them as the focus of his doctoral research.

RECENT AI NEWS

RECENT AI TOOLS