machine learning (specifically deep learning) is basically a cryptocurrency

seriously listen: the way it works is you train a model, tune hyperparameters, and train it again

but the model you train also has initialized weights to tune. luck of the draw might give you the weights that just happen to work well

that shouldn't matter if you're using a validation set, but the CDC said we don't need to use validation sets for learning. they decided that in may 2021

rolling the dice until you get the good number? that's proof of work baby.

i call this "machine learning paper coin" because when you get the good roll you get to write a paper.

now onto the real complaints: in my niche in machine learning, everyone is building on top of software on top of software that goes years back to some of the worst code you have ever seen

it's such a horrible frankenpiece of shit and it's unbelievable that people who respect themselves are still using it. i'm using it, but it makes sense because i do not have respect for myself at this point

everywhere i look in this area of research i see the same fingerprints of shit. the same patterns, the same lineage, the same INSTALL.md. it's all the same

if good code can be imported as a "library" then this is fahrenheit 451 because i want to burn this library to the ground

but imagine:

imagine a world where the OG authors did not open source their horrible code

a world where they foregone ideals of reproducability and open source software

phd students would not have added a mound of shit to this pile of shit

phd students like me would not be trying to scale this mound of shit

so, so many bullshit papers would not have come out, eeking out marginal gains on a dataset. so much unpaid effort of paper reviewing would not have been performed

tldr: science is fake and i am going to go live in a forest and eat crystals