4meta_forgetting_gpt
Year: 2025
Type: article
Abstract: I The method’s effectiveness stems from its precision: broad regularization increases train loss by 237%forcomparable speed gains, while our targeted approach limits loss increase to 8.6% through layer-wise sensitivity analysis. Thisdemonstrates that strategic parameter forgetting can enhance efficiency without compromising linguistic capability, offering a pathwayto more adaptable and maintainable language models.The method’s effectiveness stems from its precision: broad regularization increases train loss by 237%forcomparable speed gains, while our targeted approach limits loss increase to 8.6% through layer-wise sensitivity analysis. Thisdemonstrates that strategic parameter forgetting can enhance efficiency without compromising linguistic capability, offering a pathwayto more adaptable and maintainable language models.getting in anguage Models:Learning to Discard for Enhanced Efficiency and Generalizationsdf The method’s effectiveness stems from its precision: broad regularization increases train loss by 237%forcomparable speed gains, while our targeted approach limits loss increase to 8.6% through layer-wise sensitivity analysis. Thisdemonstrates that strategic parameter forgetting can enhance efficiency without compromising linguistic capability, offering a pathwayto more adaptable and maintainable language models.
Source: Figshare
Author: Huang, Zhigao
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6902236297
FWCI:
Cites:
Cited by:
Related to: