Reproduce “Regularizing Transformers With Deep Probabilistic Layers” 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
Aurora Cobo Aguilera, Pablo Martínez Olmos, Antonio Artés-Rodríguez, Fernando Pérez-Cruz
NoRBERT uses a Gaussian Mixture Variational Autoencoder (GMVAE) as a sampled hidden-state reconstruction layer inside pretrained BERT, rather than as a language generator. The authors compare insertion positions on SNLI imputation, using exact source-token matches and BLEU overlap. Early placement improves reconstruction scores; top placement yields more varied examples, but no measured diversity advantage or evidence of better completions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.