Very recommendable! A very comprehensive, impressive research paper (48 pages) by OpenAI (and authored by well known researchers like Alec Radford and Ilya Sutskever and collaborators)!
It features many evaluations on dozens of variations. "Studying 66 different models on 27 different datasets requires tuning 1782 different evaluations". The authors also tried to capture and document the historic evolution/progress of their research from early beginnings. They also created a new massive dataset of 400 million image/text pairs. What is a little surprising is that they used quite a bit of hand engineering!
Unfortunately, their research is tainted by allowing prominent leftist ideologies of the day to infiltrate it. When political, ideological indoctrination messes with the minds of researchers science is compromised! E.g. on page 24 you read following peculiar footnote "Note: The CelebA dataset is more representative of faces with
lighter skin tones. Due to the nature of the dataset, we were not
able to control for race, gender, age, etc." Had this been African researchers using an African version of CelebA, would they have noted faces with darker skin tones? Kind of absurd, definitely annoying!
No comments:
Post a Comment