www.icact.org

Member Login

TACT Journal Homepage

Call for Paper

Committee

Find My Paper

Author Homepage

Statistics & History

Welcome Message

Paper Procedure

FAQ

Paper Archives

Outstanding Papers

Program & Proceedings

Presentation Platform

Hotel & Travel Info

Photo Gallery

Scheduler Login

Seminar

Archives Login

Sponsors

Virual Oral Question and Answer


IEEE/ICACT20260070 Question.2 Questioner: Lee19711936516@163.com 2026-02-09 오후 2:57:23	IEEE/ICACT20260070 Answer.2 Answer by Auhor d11207806@mail.ntust.edu.tw 2026-02-09 오후 2:57:23
What other key designs have contributed to GIDLMMs excellent data efficiency, apart from the semantic descriptions generated by LMM and LoRA fine-tuning, despite being able to outperform CLIPping models that use 100000 samples trained on only 1000 samples?	Thank you for the question., The most critical design is that we completely freeze the pre-trained CLIP backbone. By doing this, we retain the massive general knowledge CLIP learned from 400 million images, rather than overwriting it or risking overfitting with our small dataset. We effectively force the model to use that vast prior knowledge and only learn the specific alignment between the image and our text descriptions using contrastive loss, which is far more efficient than training a classifier from scratch .

Select Voice