albert-shared-params-sop-replaces-nsp

IN premiseentries/2026/06/21/wiki-BERT_language_model.md

Created 2026-06-21T09:50:09+00:00

ALBERT (2019) shares parameters across transformer layers, replaces NSP with Sentence-Order Prediction (SOP), and decouples hidden size from embedding size

Summary

ALBERT is a BERT variant that cuts memory and training cost by reusing the same weights across all layers, swapping the Next Sentence Prediction task for a simpler sentence-shuffling task, and letting the internal representation size differ from the word-embedding size. For the system, this entry pins down the specific architectural decisions that make ALBERT lighter than BERT while staying comparably effective.

Dependents

These beliefs depend on this one: