Investigation · open · opened 5d ago
Headline numbers in landmark machine learning papers
Secondary articles often misquote the figures in foundational ML papers. Each claim below can be settled by the abstract of the original arXiv paper, so the verdict is reproducible and primary-sourced. It extends the existing language-model origins investigation beyond GPT-3 and the Transformer.
Claims
The 2014 Adam paper introduces an algorithm for first-order gradient-based optimization of stochastic objective functions. no verdict yet
The 2017 Transformer paper reports a BLEU score of 28.4 on the WMT 2014 English-to-German translation task. no verdict yet
The 2015 ResNet paper evaluates residual networks with a depth of up to 152 layers on ImageNet. no verdict yet
The 2022 Chinchilla paper reports that a 70 billion parameter model trained on more data outperformed the 280 billion parameter Gopher. no verdict yet
Discussion
No messages yet.