Paper Reading ๐Ÿ“œ

Paper Reading ๐Ÿ“œ/Natural Language Processing

Embedding Matrix ํ•™์Šต

Why I studied Embedding Matrix? NLP ๋ถ„์•ผ์—์„œ ์ค‘์š”ํ•˜๊ฒŒ ์‚ฌ์šฉ๋˜๋Š” Emebedding Matrix๋“ค์— ๋Œ€ํ•ด ํ•™์Šต์„ ํ•˜์˜€๋‹ค. Embedding Matrix๋ฅผ ํ†ตํ•ด ๋ฌธ์žฅ์—์„œ ์–ด๋– ํ•œ ๋‹จ์–ด๋‚˜ ๋ฌธ์žฅ์ด ์–ผ๋งˆ๋‚˜ ์ค‘์š”ํ•œ ์˜๋ฏธ๋ฅผ ๊ฐ€์ง€๋Š”์ง€๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋œ๋‹ค. (์ˆ˜์ •): ELMo, BERT, GPT-1์€ ๋”ฐ๋กœ ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ๋กœ ์˜ฌ๋ฆฌ๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. Table of Contents 1. What is Language Model? 2. Count based word representation 2-1. TF-iDF 3. Word Embedding 3-1. Word2Vec 3-2. GloVe 1. What is Language Model? ์–ธ์–ด ๋ชจ๋ธ (Language Model)์ด๋ž€, ์–ธ์–ด๋ฅผ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„..

Paper Reading ๐Ÿ“œ/Natural Language Processing

Better Reasoning Behind Classification Predictions with BERT for Fake News Detection ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ

Why I read this paper? fake news detection ๊ณผ์ •์—์„œ ๋ชจ๋ธ์ด ์ฃผ์–ด์ง„ ๋ฐ์ดํ„ฐ์˜ ์–ด๋– ํ•œ ๋‹จ์–ด ๋˜๋Š” ๋ฌธ์žฅ์— ๋Œ€ํ•ด ์ค‘์š”์„ฑ์„ ๋ณด์ด๋Š”์ง€ ํŒŒ์•…ํ•˜๊ธฐ ์œ„ํ•ด CAM๊ณผ Grad-CAM์„ ํ™œ์šฉํ•˜์—ฌ Visualization์„ ์ง„ํ–‰ํ•œ ๋ถ€๋ถ„์„ ์•Œ๊ธฐ ์œ„ํ•ด. Grad-CAM์€ ์ด๋ฏธ์ง€์— ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ๋œ ๊ธฐ์ˆ ์ธ๋ฐ ์ด๋ฅผ ์–ด๋–ป๊ฒŒ text data์— ์ ์šฉํ•˜๋Š”์ง€ ๊ถ๊ธˆํ•ด์„œ ์ฝ๊ฒŒ ๋˜์—ˆ๋‹ค. Table of Contents 1. Introduction 2. Methodology 3. Experiments & Results(์ผ์ • ๋ถ€๋ถ„๋งŒ) 4. Conclusion 1. Introduction ์ด ๋…ผ๋ฌธ์—์„œ๋Š” representation space์˜ ํ€„๋ฆฌํ‹ฐ๋ฅผ ์„œ๋กœ ๋‹ค๋ฅธ ๊ฐ€์งœ์™€ ์ง„์งœ ๋‰ด์Šค ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•œ linear sepa..

Paper Reading ๐Ÿ“œ/Computer Vision

Grad-CAM: Visual Explanation from Deep Networks via Gradient-based Localization ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ

Table of Contents 1. Introduction 2. Grad-CAM 1. Introduction Grad-CAM์€ Gradient-weighted Class Activation Mapping์˜ ์•ฝ์ž๋กœ, CNN์„ ํ†ตํ•ด ์ด๋ฏธ์ง€๋ฅผ ๋ถ„์„ํ•  ๋•Œ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ๋žŒ๋“ค์€ ๊ทธ ๊ณผ์ •์„ ๋ณผ ์ˆ˜ ์—†์ง€๋งŒ, Grad-CAM์„ ํ™œ์šฉํ•˜๋ฉด CNN์˜ ํ™œ๋™ ๊ณผ์ •์„ ๋”์šฑ ๋ช…๋ฐฑํ•˜๊ณ  ์ž์„ธํ•˜๊ฒŒ ์•Œ ์ˆ˜ ์žˆ๋‹ค. Grad-CAM์€ ์ด์ „์˜ ๋ชจ๋ธ๋“ค๊ณผ ๋‹ฌ๋ฆฌ ์•„๋ฌด๋Ÿฐ ๊ตฌ์กฐ์  ๋ณ€ํ™”์™€ ์žฌํ•™์Šต ์—†์ด CNN์˜ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ๋“ค์— ์ ์šฉ์ด ๊ฐ€๋Šฅํ•˜๋‹ค!! ์ด ๋…ผ๋ฌธ์—์„œ๋Š” Grad-CAM๊ณผ fine-grained visualization์„ ๊ฒฐํ•ฉํ•˜์—ฌ high-resolution class-discriminative visulaization์„ ํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ด๋ฅผ..