Hi ๐Ÿ™Œ
I am AI Traveler ๐Ÿ—บ๏ธ
who|
Hagyeong Lee.

thumbnail
Linearly Mapping from Image to Text Space

EffL LAB. Regular Seminar Linearly Mapping from Image to Text Space (ICLRโ€™23) Problem of Language Model Emily M. Bender and Alexander Koller., โ€œClimbing towards NLU: on meaning form and understanding in the age of dataโ€, ACL 2020 A System exposed only to form in its training cannot in principle learn meaning ##Form & Meaning in Language** Form Anything we can find in a language (e.g., symbols, mouth movements) Meaning Relationship between form and non-linguistic parts Including Communicativeโ€ฆ

CIPS;Image Generators with Conditionally-Independent Pixel Synthesis

์ƒ์„ฑ ๋‚œ์ด๋„๊ฐ€ ๋†’์€ ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•œ 256x256 ํ•ด์ƒ๋„์˜ ๊ฒฐ๊ณผ๋ฌผ, StyleGAN2 ๊ฒฐ๊ณผ๋ฌผ๊ณผ ์œ ์‚ฌํ•œ ์ˆ˜์ค€์„ ๋ณด์˜€๋‹ค๊ณ  ํ•œ๋‹ค. CIPS์˜ ๊ถ๊ทน์ ์ธ ๋ชฉํ‘œ๋Š” ๊ฐ ํ”ฝ์…€์„ ๋…๋ฆฝ์ ์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋ชจ๋ธ์„ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด๋‹ค. ๊ทธ๋ฅผ ์œ„ํ•ด์„œ Conv๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๋Š” ๊ฒƒ์ด ํ•„์ˆ˜์ ์ด๋ฉฐ, ๊ทธ๋Ÿผ์—๋„ ๊ณ ํ’ˆ์งˆ์˜ ์ด๋ฏธ์ง€๋ฅผ ์–ป๊ธฐ ์œ„ํ•ด Positional Encoding์„ ์ถ”๊ฐ€ํ•˜์—ฌ SoTA๋ฅผ ๋‹ฌ์„ฑํ•˜์˜€๋‹ค๋Š” ๊ฒƒ์œผ๋กœ ์š”์•ฝํ•  ์ˆ˜ ์žˆ๊ฒ ๋‹ค. Paper: https://arxiv.org/abs/2011.13775 Github: https://github.com/saic-mdal/CIPS Introduction CIPS๋Š” Spatial Convolution์ด๋‚˜ Self Attention ์—†์ด MLP๋ฅผ ์‚ฌ์šฉํ•ด ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ชจ๋ธ์ด๋‹ค. ์ผ๋ฐ˜์ ์ธ ์ƒ์„ฑ ๋ชจ๋ธ์ด Spatial Convolution์„ ์‚ฌ์šฉํ•œ ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•˜๊ณ  ์žˆ์Œ์„ ์ƒ๊ฐํ•˜๋ฉด Convolution ์—†์ด SoTA๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์ƒ๊ฐํ•  ์ˆ˜ ์—†์—ˆ์ง€๋งŒ CIPS๋Š” LSUN Churchโ€ฆ

Information Theory

Information Theory (์ •๋ณด์ด๋ก ) ์ •๋ณด์ด๋ก ์˜ ์šฉ์–ด Information : ์ •๋ณด์ด๋ก ์—์„œ๋Š” bit๋กœ ์ธก์ •๋˜๋ฉฐ ์ฃผ์–ด์ง„ ์ด๋ฒคํŠธ์—์„œ ๋ฐœ์ƒํ•˜๋Š” โ€œsurpriseโ€์˜ ์–‘์œผ๋กœ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋‹ค. (defined as the amount of โ€œsurpriseโ€ arising from a given event) ์ •๋ณด์›(Source) : ์ •๋ณด๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ๊ณณ code : ์ˆ˜์‹ ์ž๊ฐ€ ๋ฐ›์„ ์ˆ˜ ์žˆ๋Š” ๋ชจ๋“  ๋ฒกํ„ฐ๋ฅผ ์˜๋ฏธ codeword : ๋ถ€ํ˜ธ์–ด, ์ฝ”๋“œ ์ค‘์—์„œ generator๋ฅผ ํ†ตํ•ด ์ธ์ฝ”๋”ฉ๋œ ๋ฒกํ„ฐ๋งŒ์„ ์˜๋ฏธ incoding : ๋ณด๋‚ด๊ณ ์žํ•˜๋Š” ์›๋ž˜ msg(message) symbols์— ์‹๋ณ„์ž(parity check symbol)์„ ๋”ํ•˜๋Š” ๊ณผ์ • symbol : k๊ฐœ์˜ bit๋ฅผ ํ•˜๋‚˜๋กœ ๋ชจ์•„๋†“์€ ๋‹จ์œ„ bit per second (bps):์ „์†ก๋˜๋Š” bit์˜ ์ดˆ๋‹น ์†๋„ Entropy : Information์˜ ๊ธฐ๋Œ€๊ฐ’, ํŠน์ •ํ•œ stochastic process์—์„œ ์ƒ์„ฑ๋œ information์˜ ํ‰๊ท  chanโ€ฆ

Implicit Neural Representations for Image Compression

Implicit Neural Representations for Image Compression Introduction preserves all the information (lossless compression) sacrifices some information for even smaller file sizes (lossy compression) ์ •๋ณด๋ฅผ ๋ชจ๋‘ ๋ณด์กดํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ์˜ compression ๋˜๋Š” ์กฐ๊ธˆ์˜ ์ •๋ณด๋Š” ์†์‹ค์ด ์žˆ์–ด๋„ ํŒŒ์ผ ํฌ๊ธฐ๋ฅผ ๋” ์ค„์ด๋Š” ๋ฐฉํ–ฅ์œผ๋กœ์˜ compression์ด ์กด์žฌํ•œ๋‹ค. โ€”> fundamental theoretical limit (Shannonโ€™s entropy) ์ •๋ณด ์†์‹ค์—†๋Š” compression์ด ๋” desirableํ•˜์ง€๋งŒ ๊ธฐ๋ณธ ์ด๋ก ์  ํ•œ๊ณ„๊ฐ€ ์กด์žฌํ•œ๋‹ค. ์ƒค๋„Œ์˜ ์—”ํŠธ๋กœํ”ผ๋Š” ์ •๋ณด๋ฅผ ํ‘œํ˜„ํ•˜๋Š”๋ฐ ํ•„์š”ํ•œ ์ตœ์†Œ ํ‰๊ท  ์ž์›๋Ÿ‰์„ ๋งํ•˜๋Š”๋ฐ, ์ƒค๋„Œ์€ ์•„๋ฌด๋ฆฌ ์ข‹์€ ์ฝ”๋“œ๋ฅผ ์„ค๊ณ„ํ•˜๋”๋ผ๋„ ํ‰๊ท  ๊ธธ์ด๊ฐ€ ์—”ํŠธ๋กœํ”ผย H(X)๋ณด๋‹ค ์งง์•„์งˆ ์ˆ˜ ์—†์Œ์„ ๋ฐํ˜”๋‹ค. Therefore, lโ€ฆ