02. Image Classification

 정리 카테고리 Lecture 2 formalizes the problem of image classification. We discuss the inherent difficulties of image classification, and introduce data-driven approaches. We discuss two simple data-driven image classification algorithms: K-Nearest Neighbors and Linear Classifiers, and introduce the concepts of hyperparameters and cross-validation. Keywords: Image classification, K-Nearest Neighbor, distance metrics, hyperparameters, cross-validation, linear classifiers slides: http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture2.pdf # How do we work on this image classification task? - system input: images, predermined category labels - what computer really seas: just a grid  # Challenges: our classification algorithm should be robust at different kinds of transforms. 1. Viewpoint variation 1. Illumination 1. Deformation 1. Occulusion 1. Background Clutter 1. Intraclasss variation # Data-Driven Approach 1. Collect a datasset of images, labels 2. Use ML to train a classifier 3. Evaluate the classifier on new images # Classifier # K-Nearest Neighbor - majority vote among K - tends to smooth out our decision boundaries and lead to better results. - L1 distance, L2 distance - dependent on problem or data - so just reccomend to try them both and see what works better - Actually, never used 1. very slow at test time 2. Distance metrics on pixels are not informative 3. Curse of dimensionality ___ # Hyperparamets - choices about the algorithm that we set rather than learn ## Setting Hyperparameters 1. Choose hyperparameters that work best on the data (Don't do this) -> K=1 always works perfectly on training data 2. Split data into `train` and `test`; choose hyperparameters that work best on test data (Don't do this) -> No idea how algorithm will perform on new data 3. Split data into `train` and `val`, and `test`; choose hyperparameters on val and evaluate on test (Better!) 4. Cross-Validation: Split data in folds, try each fold as validation and average the results Q. training set vs validation set - algorithm doesn't have direct accesss to the labels of `validation set.` - uses validation set `only for checking` how well algorithm is doing.