This book suggests that a labeled dataset can teach a computer to experience tripping on LSD.
Training data is a collection of labeled information that is used to build a machine learning model. It can consist of annotated text, images, video, or audio. Through training data, an algorithm learns to perform a specific task, such as recognizing faces in photos or identifying cancerous lesions in x-rays. In order to develop a model to execute a new classification task, an algorithm is fed a large amount of relevant training data.
Trippy Training Data features images from a Reddit subreddit where people share their questions and experiences with taking LSD. Posts in this subreddit include people asking for advice on dosage and sharing their original psychedelic art. Some people attempt to capture their hallucinatory visions using the shoddy cameras on their smartphones. These images are often of mundane objects, such as cold pizza and bathroom floor tiles. They are snapped under drab, fluorescent lighting and captioned in all caps with expressions like, “WHAT THE FU K!”
Normally these images might belong to a collection of butt dial artifacts that get discarded to save space. However, these captions demand that we inspect the images more closely. Why were these photos worthy of such proclamations?
When an algorithm is fed labeled data, it will generalize rules, establish relationships, and detect patterns–even if there are not any meaningful ones to be found. If we stare long enough at marbled cheese, can we too experience something sublime? - Angie Waller