Nicholas Carlini publicly releases his 2016–2019 research idea log files

Nicholas Carlini published his research idea list ideas.txt, maintained from March 2016 to the end of 2019, with item-by-item retrospective notes as of January 2024. His practice was to append ideas he considered good enough at the time to a single file, reading through and filtering it only a few times a year, so that a good idea would pass at least two checks: being written down and being reread. The record focuses on the adversarial machine learning direction, such as building generators that craft minimal changes to fool a classifier, and the question of whether untargeted attacks are actually equivalent to targeted attacks aimed at the nearest other class.