<aside> đź“–
https://github.com/ml5js/Intro-ML-Arts-IMA-F24/tree/main/02-transfer-learning
</aside>
(taken from assignment 01)
Notes from the Sites:
Final Thoughts:
Computer science is - by definition - a very black and white field. Something is either a 1 or a 0, no in-between. But things get very messy when we try to apply the same logic to our daily lives and to the world we live in. It is very hard to put a single word as a label on some of the more abstract constructs in life, especially the ones which could mean different things in different contexts.
Each one us comes from a different upbringing, had different life experiences, and been exposed to different things. Thus we all experience the world vastly differently, and also views certain aspects differently. (To give a concrete example: In a country like Hungary, where gay marriage is not legal, people might not “care” as much about not having proper gender pronouns - PS: this of course is very bad stereotyping, but for the sake of making the point, I will stand by it for now.) What I’m trying to get to is we humans can’t even collectively decide on a “ground-truth” in so many areas of life, so why are we expecting a dataset to do so?!
Where we go wrong is treating these programmes as a system which is always correct (something which has been true about computers for many years; 2 + 2 is always going to be 4). But AI is not easy to navigate. I think as long as we treat these programmes as a good estimate of how most people view the world (assuming the labellers were chosen from a diverse background) it could be a useful guide, but for sure not something which tells us the one and only truth.
And yes some labels are just outright offensive, and probably should be manually overwritten. But on the other side, these tools allow for so many useful use-cases which would have never been imaginable with “conversional computing” (think of the drone surveillance example - how useful could that be at large events like festivals, which are impossible to constantly monitor)
PS: Also I’m just realising that listening to an AI version of Levitating sung by Frank Sinatra while writing this post is a bit ironic, but it’s such a good cover!