Stanfoard CS231n 2017를 요약한 포스팅입니다. 정보 전달보다 자신을 위한 정리 목적이 강한 글입니다! :)

Image Classification

• Computer Vision의 핵심 Task
  • Input : Image
  • Output : Category Labels
• Semantic Gap
  • 이미지에서 추출할 수 있는 정보(색, 질감)와 사람들이 원하는 추상적 정보의 차이
• Challenges
  • Viewpoint Variation ( 보는 각도 )
  • Illumination ( 조명 )
  • Deformation ( 변형 )
  • Occlusion ( 은폐, 숨김 )
  • Background Clutter ( 배경과 섞임 )
  • Intraclass Variation ( 물체의 다양성 )
• 상상하는 모든 이미지를 실시간으로 판단하고 싶음
• Attempts
  • Find edges
  • Find corners
  • brittle

Data-Driven Approach

• Collect a dataset of images and labels
• Use Maching Learning to train a classifier
• Evaluate the classifier on new images

Classifier : Nearest Neighbor

• Train : Memorize all data and labels
• Predict : Predict the label of the most similiar training image

Example Dataset: CIFAR10

• 10 classes(airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck)
• 50000 training images
• 10,000 testing images

Distance Metric

• L1(Manhattan) distance
• L2(Euclidean) distance

Hyper Paramter

• What is the best value of k to use?
• Whate is the best distinct to use?
• Choices about the algorithm that we set rather than learn
• Setting Hyper Paramters
  • Idea 1. Choose hyper paramters that work best on your data
    • Bad : \(K=1\) always work perfectly on training data
  • Idea 2. Split data into train and test, choose hyper parameters that work best on test data
    • Bad : No idea how algorithm will perform on new data
  • Idea 3. Split data into train, val, and test; choose hyper parameters on val and evaluate on test
    • Better!
    • validation set : check accuracy, check how well algorithm is doing
  • idea 4. Cross-Validation: Split data into folds, try each fold as validation and average the results
    • Useful for small datasets, but not used too frequently in deep learning

KNN on images never used

• Very slow at test time
• Distance metrics on pixels are not informative
• Curse of dimensionality ( 차원의 저주 )

Parametric Approach

• Image : Array of \(32\times 32\times 3\) (3072)
• Function : \(f(x,W)\)
• Output : 10 numbers giving class scores

Coming up

• Loss function : quantifying what it means to have a “good” W
• Optimization : start with random W and ifnd a W that minimizes the loss
• ConvNets : tweak the functional form of f

Reference

• Stanfoard CS231n 2017

카일스쿨 유튜브 채널을 만들었습니다. 데이터 사이언스, 성장, 리더십, BigQuery 등을 이야기할 예정이니, 관심 있으시면 구독 부탁드립니다 :)

PM을 위한 데이터 리터러시 강의를 만들었습니다. 문제 정의, 지표, 실험 설계, 문화 만들기, 로그 설계, 회고 등을 담은 강의입니다

이 글이 도움이 되셨거나 다양한 의견이 있다면 댓글 부탁드립니다 :)