no image
[Core Review] OOD Detection with Relative Angles
๋…ผ๋ฌธ ๋งํฌ : ORAOOD Detection with Relative Angles ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ์ž…๋‹ˆ๋‹ค.Motivation๊ธฐ์กด์˜ post-hoc OOD detection ์—ฐ๊ตฌ๋“ค์€ ํฌ๊ฒŒ Logit ๊ธฐ๋ฐ˜๊ณผ Distance ๊ธฐ๋ฐ˜์œผ๋กœ ๋‚˜๋‰œ๋‹ค. ํ•˜์ง€๋งŒ ์ €์ž๋“ค์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ฌธ์ œ์ ์„ ์ง€์ ํ•œ๋‹ค:๊ธฐ์กด ๊ฑฐ๋ฆฌ ๊ธฐ๋ฐ˜ ๋ฐฉ์‹์˜ ๋งน์ : KNN์ด๋‚˜ Mahalanobis ๊ฑฐ๋ฆฌ ๋“ฑ์€ feature space์—์„œ์˜ ๊ฑฐ๋ฆฌ๋ฅผ ์ธก์ •ํ•˜์ง€๋งŒ, In-distribution(ID) ๋ฐ์ดํ„ฐ์˜ ํ†ต๊ณ„์  ๊ตฌ์กฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋ฐ˜์˜ํ•˜์ง€ ๋ชปํ•˜๊ฑฐ๋‚˜ decision boundary์™€์˜ ๊ด€๊ณ„๋ฅผ ๊ฐ„๊ณผํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค.fDBD(Fast Decision Boundary-based Detector)์˜ ํ•œ๊ณ„: ์ตœ๊ทผ decision boundary๊นŒ์ง€์˜ ๊ฑฐ๋ฆฌ๋ฅผ ํ™œ์šฉํ•˜๋Š” fDBD๊ฐ€ ์ œ..
2026.03.19
no image
[Core Review] Optimal Parameter and Neuron Pruning for Out-of-Distribution Detection
๋…ผ๋ฌธ ๋งํฌ : OPNPOptimal Parameter and Neuron Pruning for Out-of-Distribution Detection ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ์ž…๋‹ˆ๋‹ค.Motivationํ˜„๋Œ€์˜ ์‹ ๊ฒฝ๋ง์€ OOD sample์— ๋Œ€ํ•ด overconfidentํ•œ ๊ฒฝํ–ฅ์ด ์žˆ์–ด ID data์™€ OOD data์˜ ๊ตฌ๋ถ„์„ ์–ด๋ ต๊ฒŒ ๋งŒ๋“ ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์˜ Training-based method๋Š” OOD sample์ด ํ•™์Šต ์‹œ ํ•„์š”ํ•˜๋ฉฐ ์—ฐ์‚ฐ ๋น„์šฉ์ด ๋†’๋‹ค. Post-hoc method๋Š” ์ถ”๊ฐ€ ํ•™์Šต์ด ํ•„์š” ์—†๊ณ  ๊ฐ„ํŽธํ•˜์ง€๋งŒ, ํ•™์Šต data์— ํฌํ•จ๋œ Prior information์„ ์ถฉ๋ถ„ํžˆ ํ™œ์šฉํ•˜์ง€ ๋ชปํ•œ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋งˆ์ง€๋ง‰ Fully-connected layer์˜ parameter Sensitivity๋ฅผ ๋ถ„์„ํ•˜์—ฌ Sensitivi..
2026.02.15
no image
About & Privacy
[๋ธ”๋กœ๊ทธ ์†Œ๊ฐœ]์•ˆ๋…•ํ•˜์„ธ์š”! ๋Š์ž„์—†์ด ๋ฐœ์ „ํ•˜๋Š” AI ๊ธฐ์ˆ ๊ณผ ๋”ฅ๋Ÿฌ๋‹(Deep Learning)์˜ ์„ธ๊ณ„๋ฅผ ํƒ๊ตฌํ•˜๋Š” ๋ธ”๋กœ๊ทธ์ž…๋‹ˆ๋‹ค.์ด ๋ธ”๋กœ๊ทธ๋Š” ์ œ๊ฐ€ ์ง์ ‘ ์ฝ๊ณ  ๋ถ„์„ํ•œ ์ตœ์‹  AI/ML ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ(Paper Reviews)์™€ ํ•™์Šต ๊ณผ์ •์„ ๊ธฐ๋กํ•˜๋Š” ์•„์นด์ด๋ธŒ์ž…๋‹ˆ๋‹ค. ๋ณต์žกํ•œ ์ˆ˜์‹๊ณผ ์ด๋ก ์„ ์ดํ•ดํ•˜๊ธฐ ์‰ฝ๊ฒŒ ์ •๋ฆฌํ•˜์—ฌ ์ง€์‹์„ ๊ณต์œ ํ•˜๊ณ , ํ•จ๊ป˜ ์„ฑ์žฅํ•˜๋Š” ๊ณต๊ฐ„์ด ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.์ฃผ์š” ๋‹ค๋ฃจ๋Š” ์ฃผ์ œ:Paper Reviews: Computer Vision, NLP ๋“ฑ ์ฃผ์š” ํ•™ํšŒ(CVPR, NeurIPS ๋“ฑ) ๋…ผ๋ฌธ ์‹ฌ์ธต ๋ถ„์„Deep Learning Concepts: ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ์ดˆ ์ด๋ก  ๋ฐ ๊ตฌํ˜„Tech Tips: ๊ฐœ๋ฐœ ํ™˜๊ฒฝ ์„ธํŒ… ๋ฐ Mac ํ™œ์šฉ ํŒ๋ฐฉ๋ฌธํ•ด ์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋‚ด์šฉ์— ์˜ค๋ฅ˜๊ฐ€ ์žˆ๊ฑฐ๋‚˜ ํ† ๋ก ํ•˜๊ณ  ์‹ถ์€ ์ฃผ์ œ๊ฐ€ ์žˆ๋‹ค๋ฉด ์–ธ์ œ๋“  ๋Œ“๊ธ€์ด๋‚˜ ๋ฉ”์ผ๋กœ..
2026.02.10
no image
[Paper Review] Out-of-Distribution Detection with Negative Prompts
๋…ผ๋ฌธ ๋งํฌ : LSNOut-of-Distribution Detection with Negative Prompts ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ์ž…๋‹ˆ๋‹ค.Introduction๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์€ ๋Œ€๊ทœ๋ชจ ๋ผ๋ฒจ ๋ฐ์ดํ„ฐ๊ฐ€ ์ฃผ์–ด์งˆ ๋•Œ ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€๋‹ค. ํ•˜์ง€๋งŒ ๋Œ€๋ถ€๋ถ„์˜ ๋ฐฉ๋ฒ•์€ ํ…Œ์ŠคํŠธ ์‹œ์ ์˜ ํด๋ž˜์Šค๊ฐ€ ํ•™์Šต ์‹œ์ ์˜ ํด๋ž˜์Šค์™€ ๋™์ผํ•˜๋‹ค๋Š” ์ „์ œ์ธ closed-set ๊ฐ€์ •์— ๊ธฐ๋ฐ˜ํ•œ๋‹ค. ๋ฌธ์ œ๋Š”, ์‹ค์ œ ํ™˜๊ฒฝ์—์„œ๋Š” ํ•™์Šต ์ค‘ ๋ณด์ง€ ๋ชปํ–ˆ๋˜ OOD data๊ฐ€ ์–ธ์ œ๋“  ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ๋‹ค. ์ด๋•Œ, ๋ชจ๋ธ์€ ID data๋„ ์ •ํ™•ํ•˜๊ฒŒ, OOD data๋„ ์ •ํ™•ํ•˜๊ฒŒ ๋ถ„๋ฅ˜ํ•ด์•ผํ•œ๋‹ค. ๊ธฐ์กด์˜ ๋งŽ์€ OOD detection ๋ฐฉ๋ฒ•์€ post-hoc ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•œ๋‹ค. ์ด ๋ฐฉ๋ฒ•์˜ ๋ฌธ์ œ์ ์€ ๋ถ„๋ฅ˜์— ์œ ์šฉํ•œ feature๊ฐ€ OOD detection์—๋„ ์œ ์šฉํ•˜๋‹ค๋Š” ๋ณด์žฅ์ด ์—†๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ์˜ˆ๋ฅผ..
2026.02.10
no image
[Paper Review] Out-of-Distribution Detection by Leveraging Between-Layer Transformation Smoothness
๋…ผ๋ฌธ ๋งํฌ : BLOODOut-of-Distribution Detection by Leveraging Between-Layer Transformation Smoothness ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ ์ž…๋‹ˆ๋‹คIntroductionMachine Learning model์˜ ์„ฑ๊ณต์€ ํ•™์Šต data์™€ ๋™์ผํ•œ ๋ถ„ํฌ์—์„œ ์ƒ์„ฑ๋œ in-distribution (ID) data์—์„œ ํ‰๊ฐ€๋œ๋‹ค๋Š” ๊ฐ€์ •์— ๊ธฐ๋ฐ˜ํ•œ๋‹ค. ํ•˜์ง€๋งŒ noise๊ฐ€ ์กด์žฌํ•˜๊ณ  ๋ถˆ์™„์ „ํ•œ ์‹ค์ œ ํ™˜๊ฒฝ์— ๋ฐฐํฌ๋œ model์€ ์ข…์ข… ์„œ๋กœ ๋‹ค๋ฅธ ๋ถ„ํฌ์—์„œ ์ƒ์„ฑ๋œ Out-of-distribution (OOD) data์— ์ง๋ฉดํ•˜๊ฒŒ ๋˜๋ฉฐ, ์ด๋Š” model ์„ฑ๋Šฅ์„ ์ €ํ•˜์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค. ๋‹ค์–‘ํ•œ Task์—์„œ ์ผ๊ด€๋˜๊ฒŒ SOTA ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” DNN์€ OOD detection์—ฐ๊ตฌ์—์„œ ํฐ ์ฃผ๋ชฉ์„ ๋ฐ›์•„์™”๋‹ค. ..
2026.02.01
no image
[Core Review] A Critical Evaluation of Open-World Machine Learning
๋…ผ๋ฌธ ๋งํฌ : A Critical Evaluation of Open-World Machine LearningA Critical Evaluation of Open-World Machine Learning ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ์ž…๋‹ˆ๋‹ค.Motivation๊ธฐ์กด Open-World ML ์—ฐ๊ตฌ๋“ค์€ ID data๋กœ ํ•™์Šต๋œ Closed-world model์— OOD detector๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ, ํ•™์Šต๋˜์ง€ ์•Š์€ ๋ถ„ํฌ์˜ ๋ฐ์ดํ„ฐ๋ฅผ rejectํ•˜๋Š” ๋ฐฉ์‹์„ ์ทจํ•œ๋‹ค.์ œํ•œ์ ์ธ ํ‰๊ฐ€ ํ™˜๊ฒฝ: ๊ธฐ์กด ์—ฐ๊ตฌ๋“ค์€ ๊ณ ์ •๋œ ID data, model architecture, ํ•œ์ •๋œ OOD dataset๋งŒ์„ ์‚ฌ์šฉํ•˜์—ฌ ์„ฑ๋Šฅ์„ ์ธก์ •Input data๊ฐ€ clean ์ƒํƒœ๋ผ๊ณ  ๊ฐ€์ •ํ•˜๋ฉฐ, data์˜ Corruption์ด๋‚˜ Adversarial Perturbation์ด ์กด..
2026.01.30
no image
[MacOS] ๋งฅ ๋“œ๋ž˜๊ทธ ์•ค ๋“œ๋ž ์•ˆ๋˜๋Š” ๊ฒฝ์šฐ
๋งฅ๋ถ Finder ๋“œ๋ž˜๊ทธ ์•ค ๋“œ๋กญ ๋จนํ†ต ํ˜„์ƒ ํ•ด๊ฒฐ ๋ฐฉ๋ฒ•๋งฅ(Mac)์„ ์‚ฌ์šฉํ•˜๋‹ค ๋ณด๋ฉด ๊ฐ‘์ž๊ธฐ ํŒŒ์ผ์ด๋‚˜ ํด๋”๊ฐ€ ๋“œ๋ž˜๊ทธ๋˜์ง€ ์•Š๋Š” '๋“œ๋ž˜๊ทธ ์•ค ๋“œ๋กญ(Drag & Drop) ๋ฒ„๊ทธ'๋ฅผ ๊ฒฝํ—˜ํ•  ๋•Œ๊ฐ€ ์žˆ๋‹ค. ํŒŒ์ผ ์ด๋™์ด ๋ถˆ๊ฐ€๋Šฅํ•ด์ง€๋‹ˆ ๋งค์šฐ ๋‹ต๋‹ตํ•œ ์ƒํ™ฉ์ธ๋ฐ, ์ด๋Š” ์ฃผ๋กœ macOS์˜ Finder ํ”„๋กœ์„ธ์Šค๊ฐ€ ์ผ์‹œ์ ์œผ๋กœ ์ถฉ๋Œํ•˜๊ฑฐ๋‚˜ ๊ผฌ์˜€์„ ๋•Œ ๋ฐœ์ƒํ•œ๋‹ค.์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๊ฐ€์žฅ ํ™•์‹คํ•˜๊ณ  ๋น ๋ฅธ ๋ฐฉ๋ฒ• 2๊ฐ€์ง€๋ฅผ ์†Œ๊ฐœํ•œ๋‹ค.๋ฐฉ๋ฒ• 1: ํ„ฐ๋ฏธ๋„(Terminal)์„ ์ด์šฉํ•œ Finder ์žฌ์‹คํ–‰๊ฐ€์žฅ ๋น ๋ฅด๊ณ  ๊น”๋”ํ•œ ๋ฐฉ๋ฒ•์€ ํ„ฐ๋ฏธ๋„ ๋ช…๋ น์–ด๋ฅผ ํ†ตํ•ด Finder๋ฅผ ๊ฐ•์ œ๋กœ ์žฌ์‹œ์ž‘ํ•˜๋Š” ๊ฒƒ.Finder๋Š” ์ข…๋ฃŒ๋˜์–ด๋„ OS์— ์˜ํ•ด ์ฆ‰์‹œ ์ž๋™์œผ๋กœ ๋‹ค์‹œ ์‹คํ–‰๋˜๋ฏ€๋กœ ๊ฑฑ์ •ํ•˜์ง€ ์•Š์œผ์…”๋„ ๋ฉ๋‹ˆ๋‹ค.Command + Space๋ฅผ ๋ˆŒ๋Ÿฌ ์ŠคํฌํŠธ๋ผ์ดํŠธ ๊ฒ€์ƒ‰์„ ์—ฝ๋‹ˆ๋‹ค.Terminal (๋˜..
2026.01.26
no image
[Paper Revew] HaloScope: Harnessing Unlabeled LLM Generations for Hallucination Detection
๋…ผ๋ฌธ ๋งํฌ : HaloScopeHaloScope: Harnessing Unlabeled LLM Generations for Hallucination Detection ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ์ž…๋‹ˆ๋‹ค.Introduction์˜ค๋Š˜๋‚  ๋น ๋ฅด๊ฒŒ ์ง„ํ™”ํ•˜๋Š” machine learningํ™˜๊ฒฝ์—์„œ, LLMs์€ ๋‹ค์–‘ํ•œ ์‘์šฉ ๋ถ„์•ผ๋ฅผ ํ˜•์„ฑํ•˜๋Š” ๊ธฐ์ˆ ๋กœ ๋– ์˜ค๋ฅด๊ณ  ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์„ฑ๋Šฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ , Open-worldํ™˜๊ฒฝ์— ๋ฐฐํฌ๋  ๊ฒฝ์šฐ ๋ชจ๋ธ์˜ ์‹ ๋ขฐ์„ฑ๊ณผ ๊ด€๋ จ๋œ ์—ฌ๋Ÿฌ ๋„์ „ ๊ณผ์ œ๊ฐ€ ์กด์žฌํ•œ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, LLM์€ ์ธ๊ฐ„๊ณผ์˜ ์ƒํ˜ธ์ž‘์šฉ ๊ณผ์ •์—์„œ ๊ฒ‰๋ณด๊ธฐ์—๋Š” ์œ ์šฉํ•ด ๋ณด์ด์ง€๋งŒ ์‚ฌ์‹ค๊ณผ ๋‹ค๋ฅธ ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ์ค‘์š”ํ•œ ์˜์‚ฌ๊ฒฐ์ •์„ ์œ„ํ—˜์— ๋น ๋œจ๋ฆด ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” LLM์€ ํ”„๋กฌํ”„ํŠธ์™€ ์ผ๊ด€๋œ ํ…์ŠคํŠธ๋ฅผ ์ •ํ™•ํžˆ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ, hal..
2026.01.20
no image
[Core Review] Learning to Shape In-distribution Feature Space for Out-of-distribution Detection
๋…ผ๋ฌธ ๋งํฌ : DRLLearning to Shape In-distribution Feature Space for Out-of-distribution Detection ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ์ž…๋‹ˆ๋‹ค.Motivation๊ธฐ์กด์˜ OOD detection ๋ฐฉ๋ฒ•๋“ค, ํŠนํžˆ density-based ๋ฐ post-hoc method๋“ค์€ ์•„๋ž˜์™€ ๊ฐ™์€ ๊ทผ๋ณธ์ ์ธ ํ•œ๊ณ„์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค.Distributional Mismatch: ๊ธฐ์กด ๋ฐฉ๋ฒ•๋“ค์€ ํ•™์Šต๋œ feature space์— ๋Œ€ํ•ด GMM์ด๋‚˜ Gibbs-Boltzmann ๋ถ„ํฌ ๊ฐ™์€ ํŠน์ • ๋ถ„ํฌ๋ฅผ ๊ฐ€์งˆ ๊ฒƒ์ด๋ผ๊ณ  ์‚ฌํ›„์ ์œผ๋กœ ๊ฐ€์ •ํ•œ๋‹ค. ํ•˜์ง€๋งŒ ์‹ค์ œ ํ•™์Šต๋œ ํŠน์ง•๋“ค์€ ์ด๋Ÿฌํ•œ ๊ฐ€์ •์„ ๋”ฐ๋ฅด์ง€ ์•Š๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•„ ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ๋ฐœ์ƒํ•œ๋‹ค.์ผ๋ฐ˜์ ์ธ CE loss๋ฅผ ํ™œ์šฉํ•œ ํ•™์Šต์€ classification์„ฑ๋Šฅ์„ ์ตœ์ ..
2026.01.07