[Paper Review] CLIP Surgery for Better Explainability with Enhancement in Open-Vocabulary Tasks

Автор: 서울대학교 산업공학과 DSBA 연구실

Загружено: 2024-01-11

Просмотров: 4032

Описание:

발표자: 고려대학교 DSBA 연구실 석사과정 남웅찬([email protected])

1. 논문 제목 : CLIP Surgery for Better Explainability with Enhancement in Open-Vocabulary Tasks (arXiv preprint, 2023)

2. 원문 링크 : https://arxiv.org/abs/2304.05653

3. 요약
CLIP의 Explainability와 관련하여 2가지 문제점을 지적
Foreground에 집중하지 못하고 Background에 집중하는 Opposite Visualization 문제점을 기존의 Self-Attention Block 내의 Query-Key 연산을 Value-Value 연산으로 바꾼 후 이를 New Block으로 정의한 후 Original Self-Attention Block과 함께 Dual Path로 구성하는 방법을 제안
Attention Map내에서 Uniform하게 Noise 형태로 Activation이 되는 Noisy Activation 문제점을 feature 내의 불필요한 Redundant features를 계산하여 기존의 Multiplied features에서 빼는 Feature Surgery 전략을 제안

4. Keyword : #CLIP Surgery, #CLIP, #Explainability

[Paper Review] CLIP Surgery for Better Explainability with Enhancement in Open-Vocabulary Tasks

Доступные форматы для скачивания:

Скачать видео mp4

Информация по загрузке:

Скачать аудио mp3

Похожие видео

[Paper Review] BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language

[Paper Review] BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language

[Paper Review] Memory bank-guided diffusion model for lightweight anomaly detection

[Paper Review] Memory bank-guided diffusion model for lightweight anomaly detection

Why "Simple" Bidders Make Auctions Harder: The Complexity Paradox

[Paper Review] An Inverse Scaling Law for CLIP Training

[Paper Review] An Inverse Scaling Law for CLIP Training

[DSBA] Lab Seminar 2024

[DSBA] Lab Seminar 2024

[Paper Review] Parameter-Inverted Image Pyramid Networks

[Paper Review] Parameter-Inverted Image Pyramid Networks

[Paper Review] Training LLMs to be Better Text Embedders through Bidirectional Reconstruction

[Paper Review] Training LLMs to be Better Text Embedders through Bidirectional Reconstruction

Czy Chiny szykują się na upadek Rosji?

Czy Chiny szykują się na upadek Rosji?

[Paper Review] Patch-wise Structural Loss for Time Series Forecasting

[Paper Review] Patch-wise Structural Loss for Time Series Forecasting

Tragedia na zamarzniętym stawie. Nie żyje 7-latek, jego matka walczy o życie

Tragedia na zamarzniętym stawie. Nie żyje 7-latek, jego matka walczy o życie

[Paper Review] Post-pre-training for Modality Alignment in Vision-Language Foundation Models

[Paper Review] Post-pre-training for Modality Alignment in Vision-Language Foundation Models

KAIST XAI Tutorial 2025 | Fast and Faithful Explanations | Youngjin Park (SAIL, KAIST)

KAIST XAI Tutorial 2025 | Fast and Faithful Explanations | Youngjin Park (SAIL, KAIST)

[Paper Review] TAIL MIL

[Paper Review] TAIL MIL

[Paper Review] Solving the Cold-Start Problem in Short-Term Load Forecasting Using Tree-Based Method

[Paper Review] Solving the Cold-Start Problem in Short-Term Load Forecasting Using Tree-Based Method

Całą noc w aucie. Dramat kierowców na S7

Całą noc w aucie. Dramat kierowców na S7

[김종학의 뉴욕, 지금-12월24일] 트럼프

[김종학의 뉴욕, 지금-12월24일] 트럼프 "차기 의장, 호황 때 금리 내려야" | 헤셋 “일자리 급증할 것" | 종목(비자, 마스터카드, 화이자, 록히드마틴)

'문명'이 '문자'보다 앞설 수 있었던 이유| 주원준 한님성서연구소 수석연구원 [신과대화]

'문명'이 '문자'보다 앞설 수 있었던 이유| 주원준 한님성서연구소 수석연구원 [신과대화]

#666 Rosja wybrała cele, Ataki na Wenezuelę, Konflikt KSA-UAE, Protesty w Iranie, Wybory w Kosowie

#666 Rosja wybrała cele, Ataki na Wenezuelę, Konflikt KSA-UAE, Protesty w Iranie, Wybory w Kosowie

PILNA NARADA Z PREZYDENTEM - SYTUACJA JEST POWAŻNA

PILNA NARADA Z PREZYDENTEM - SYTUACJA JEST POWAŻNA

Nawrocki ujawnia „zdradę” Tuska? Uderzające nagranie wywołuje burzę

Nawrocki ujawnia „zdradę” Tuska? Uderzające nagranie wywołuje burzę