This repo contains a classifier on Fashion-MNIST dataset (https://github.com/zalandoresearch/fashion-mnist). We attempt to build the classigifer with highest accuracy while using the smallest number of training label possible.
We levarage the power of weak supervision (https://www.snorkel.org/blog/weak-supervision) to generate large number of labels by using labelling functions built from a small subset of labeled dataset.
- KNN on raw pixel
- KNN on HOG feature extraction
- SVM on HOG feature extraction
- Small CNN network due to a limitation in computation resources and time.