Reproduce “The Unreasonable Effectiveness of Random Target Embeddings for Continuous-Output Neural Machine Translation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Evgeniia Tokarchuk, Vlad Niculae
The authors test whether continuous-output neural machine translation needs meaningful target vectors by replacing pretrained vectors with fixed random ones. The decoder predicts a vector for each token. Random vectors are close to pretrained vectors overall, but results vary: on WMT 2019 English-to-German they score 31.8 BLEU versus 29.2, while underperforming on lower-resource English-to-Turkish.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.