Social Media Images: Understanding Food Perception with Deep Learning

This lecture delves into the use of Instagram images and hashtags for deep learning in understanding food perception, obesity rates, food deserts, depression, and mental health. It explores weakly supervised pretraining with 1B+ Instagram images, showcasing the shift from supervised to weakly supervised training. The lecture discusses the SEER self-supervised learning model, which outperformed state-of-the-art systems after pretraining on a billion Instagram images. It also covers the affordances of social media, such as persistence, visibility, and spreadability, and the concept of ephemeral social media like Snapchat. Computer vision plays a crucial role in analyzing social images, with attention being drawn to images containing faces. The lecture concludes with insights on social images as a form of community activity and the implications of ephemerality in design.