Pith. sign in

Speeding up Deep Model Training by Sharing Weights and Then Unsharing

1 Pith paper cite this work. Polarity classification is still indexing.

1 Pith paper citing it
abstract

We propose a simple and efficient approach for training the BERT model. Our approach exploits the special structure of BERT that contains a stack of repeated modules (i.e., transformer encoders). Our proposed approach first trains BERT with the weights shared across all the repeated modules till some point. This is for learning the commonly shared component of weights across all repeated layers. We then stop weight sharing and continue training until convergence. We present theoretic insights for training by sharing weights then unsharing with analysis for simplified models. Empirical experiments on the BERT model show that our method yields better performance of trained models, and significantly reduces the number of training iterations.

fields

cs.LG 1

years

2025 1

verdicts

CONDITIONAL 1

representative citing papers

Parallel Scaling Law for Language Models

cs.LG · 2025-05-15 · conditional · novelty 6.0

Running P parallel streams of a language model with learned prefixes gives performance comparable to multiplying parameters by about k log P + 1, with k around 0.33 to 0.39.

citing papers explorer

Showing 1 of 1 citing paper.

  • Parallel Scaling Law for Language Models cs.LG · 2025-05-15 · conditional · none · ref 95 · internal anchor

    Running P parallel streams of a language model with learned prefixes gives performance comparable to multiplying parameters by about k log P + 1, with k around 0.33 to 0.39.