๐Ÿ“ AI & Bigdata/Paper Review

๐Ÿ“ AI & Bigdata/Paper Review

[๋…ผ๋ฌธ๋ฆฌ๋ทฐ] XGBoost: A scalable Tree Boosting System

๋ฐ์ดํ„ฐ ๋ถˆ๊ท ํ˜• ์ฒ˜๋ฆฌ (SMOTE) ๋ฐ์ดํ„ฐ ๋ถˆ๊ท ํ˜• ์ฒ˜๋ฆฌ๋ฒ• ์ด๋ฒˆ์—๋Š” ๋ฐ์ดํ„ฐ ๋ถˆ๊ท ํ˜•์— ๋Œ€ํ•ด ํ•œ๋ฒˆ ์งš์–ด๋ณด๊ณ  ๊ฐ€๊ฒ ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถˆ๊ท ํ˜•ํ•œ ๊ฒฝ์šฐ ... blog.naver.com boosting: ์†๋„๊ฐ€ ๋Š๋ฆฌ๊ณ  overfitting์ด ์‰ฝ๋‹ค. bagging์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ์ข‹๊ณ  ํ•ด์„์ด ์‰ฝ๋‹ค. AdaBoost & GBM(Gradient Boost) AdaBoost: GBM๊ณผ ์œ ์‚ฌํ•œ ๋งค์ปค๋‹ˆ์ฆ˜. ๋‹ค์Œ์— ์˜ค๋ถ„๋ฅ˜๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒ๋  ํ™•๋ฅ ์ด ๋†’์•„์ง€๋Š” ๊ฒƒ GBM: ๊ฒฝ์‚ฌํ•˜๊ฐ•๋ฒ•์„ ์ด์šฉํ•˜์—ฌ ์ƒˆ๋กœ ๊ฐ€์ค‘์น˜๋ฅผ ์—…๋ฐ์ดํŠธํ•œ๋‹ค. ์ด์ „ ๋ชจ๋ธ์˜ ์ž”์ฐจ๋ฅผ ๊ฐ€์ง€๊ณ  weak model์„ ๊ฐ•ํ™”์‹œํ‚จ๋‹ค.(nagative gradient) ์˜ค๋ถ„๋ฅ˜๊ฐ’์„ ์ด์šฉํ•˜๋Š” AdaBoost์™€ ๋‹ค๋ฅด๊ฒŒ Gradient Boost์€ ๊ธฐ์šธ๊ธฐ๋ฅผ ์ด์šฉํ•œ๋‹ค. ์˜ˆ์ธก ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ๋‹ค. ํ•™์Šต ์‹œ๊ฐ„..

๐Ÿ“ AI & Bigdata/Paper Review

[๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,Shaoqing Ren et.al., NIPS 2015

Fast R-CNN notable drawback Fast rcnn์€ Selective Search๊ฐ€ ๋…๋ฆฝ์ ์œผ๋กœ ์ˆ˜ํ–‰๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์— bottleneck์ด ๋ฐœ์ƒํ•˜์˜€์Šต๋‹ˆ๋‹ค.(region proposal์€ CPU์—ฐ์‚ฐ์ด๊ณ  region-based CNN์€ GPU ์—ฐ์‚ฐ) ๊ทธ๋ž˜์„œ detection ๋„คํŠธ์›Œํฌ์˜ ์„ฑ๋Šฅ์„ ์•„๋ฌด๋ฆฌ ๊ฐœ์„ ์‹œ์ผœ ๋ดค์ž Selective Search, region proposals์˜ ์‹œ๊ฐ„์€ ๊ทธ๋Œ€๋กœ ์ถ”๊ฐ€๊ฐ€ ๋˜๊ธฐ ๋•Œ๋ฌธ์— ์†๋„๋ฅผ ๊ฐœ์„ ์‹œํ‚ค์ง€ ๋ชปํ•œ๋‹ค๋Š” ๋‹จ์ ์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์‹ค์‹œ๊ฐ„ ์ถ”๋ก ์„ ๋ชฉํ‘œ๋กœ ์ œ์•ˆ ํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ๋งค์šฐ ๋น ๋ฅธ ์ถ”๋ก ์ด ๊ฐ€๋Šฅํ•˜๋‹ค 1. Faster R-CNN: RPN + Fast R-CNN ๊ทธ๋ž˜์„œ faster rcnn์€ Cpu์—์„œ ์ง„ํ–‰๋˜๋Š” Selective search์˜ ๋ฐฉ์‹์ด ์•„๋‹Œ Region propo..

๐Ÿ“ AI & Bigdata/Paper Review

[๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] Fast R-CNN: Ross Girshick, ICCV 2015

R-CNN notable drawback 1.Training is a multi-stage pipeline 2.The image was forcibly warped to a size of 224x224 3.No back propagation 4.Training is expensive in space and time(disk, 2.5 GPU-days) 2000๊ฐœ์˜ Image Proposal ํ›„๋ณด๋ฅผ ๋ชจ๋‘ CNN ๋ชจ๋ธ์— ์ง‘์–ด ๋„ฃ๊ธฐ ๋•Œ๋ฌธ์—, training, testing ์‹œ๊ฐ„์ด ๋งค์šฐ ์˜ค๋ž˜ ๊ฑธ๋ฆผ. AlexNet์„ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด Image๋ฅผ 224x224 ํฌ๊ธฐ๋กœ ๊ฐ•์ œ๋กœ warping ์‹œ์ผฐ๊ธฐ ๋•Œ๋ฌธ์— ์ด๋ฏธ์ง€ ๋ณ€ํ˜•์œผ๋กœ ์ธํ•œ ์„ฑ๋Šฅ ์†์‹ค์ด ์กด์žฌ ๋’ท ๋ถ€๋ถ„์—์„œ ์ˆ˜ํ–‰ํ•œ Computation์„ Shareํ•˜์ง€ ์•Š๋Š”๋‹ค. (N..

๐Ÿ“ AI & Bigdata/Paper Review

[๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] R-CNN: Rich feature hierarchies for accurate object detection and semantic segmentation CVPR 2013

๋…ผ๋ฌธ: Rich feature hierarchies for accurate object detection and semantic segmentation, Ross Girshick et al. CVPR. 2013 Referance: https://arxiv.org/pdf/1311.2524.pdf ๊ถ๊ธˆํ•œ ์ ์€ ๋Œ“๊ธ€๋กœ ๋‚จ๊ฒจ์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ๐Ÿ˜Š ์ž ๊น!! Object Detection์ด๋ž€.. - input: image - ouput: bounding box(Regression), Classification Regression(bouding box์˜ ์ขŒํ‘œ๊ฐ’)์„ ํ†ตํ•ด object๊ฐ€ ์–ด๋””์— ์žˆ๋Š”์ง€, object๊ฐ€ ๋ฌด์—‡์ธ์ง€(Classification) ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์ด Object Detection์ด๋‹ค. 2013.11..

SOIT
'๐Ÿ“ AI & Bigdata/Paper Review' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๊ธ€ ๋ชฉ๋ก