Large Memory Layers with Product Keys
Augmenting transformer language models with sparse access of large memory matrices
Pattern-Exploiting Training
Leveraging the knowledge locked away in language models by reframing categorical tasks as constrained text generation.
Optimal Transport and the Sinkhorn Transformer
Optimal Transport, the Sinkhorn Transformer, and Charmin Ultra-Soft
A Survey of Methods for Model Compression in NLP
A foray into numeric precision reduction, operation fusion, pruning, knowledge distillation, and module replacement.
Talking-Heads Attention
Put on your headphones, jam out to some funky 80s rock and read about an equally funky variation on multi-head attention.
Finetuning Transformers with JAX + Haiku
A practical, code-first look at DeepMind's new haiku library.
A Survey of Long-Term Context in Transformers
Exploring 6 noteworthy approaches for incorporating longer-term context in transformer models.
Rationale Informed Classification
Leveraging annotator rationales for more interpretable and sample efficient classification.
machine learning musings