In this project, we investigate the mainstream techniques used in hand-written digit recognition task and propose semi-supervise(self-supervise) approaches in dealing with digits recognition with only a small fraction of labeled images. The dataset given for this project is based on Mixed National Institute of Standards and Technology database (MNIST) with transformations like adding noise and moving digit center. Traditional machine learning models like softmax regression, support vector machine (SVM), k-nearest neighbors(k-NN) as well as more advanced deep learning approaches like VGG, Resnet are also included. Some models can achieve nearly perfect performance(99%+ accuracy). Finally, we present our new method which shows comparable performance to most state-of-art models while only use negligible samples.

1. Download and extract the dataset:

   ./dataset.sh

2. Traditional Machine Learning:

   You can run SVM model just like below. The model_type can be svm, nusvm, knn, sr. And you can see more optioal with python main.py --help

   cd traditional_machine_learning/
   python main.py --model_type svm
   # if you want to use PCA and knn
   python main.py --model_type knn --pca

3. ResNet and VGG:

   cd resnet_and_vgg/
   # if you want to try vgg
   python main_vgg.py
   # if you want to try resnet
   python main_resnet.py

4. Virtual Adersarial Training

   You can see more details about VAT in ./virtual_adversarial_training/README.md.

   cd virtual_adversarial_training
   python train_sup.py --cost_type=VAT_finite_diff --epsilon=2.1 --layer_sizes=784-1200-600-300-150-10 --save_filename=<filename>
   python train_semisup.py --cost_type=VAT_finite_diff --epsilon=0.3 --layer_sizes=784-1200-1200-10 --num_labeled_samples=100 --save_filename=<filename>

5. Deep Generate Model:

   cd deep_generate_model/
   python main.py

1. Traditional Machine Learning

   Model	Training Accuracy	Testing Accuracy
   Softmax Regression	32	29
   KNN	-	89
   SVM	94	93
   MLP	92	87

2. Deep Learing Method

   TrainSet Size	VGG	ResNet	DGM	VAT
   100	50.74	51.52	90.09	98.07
   1000	92.23	95.81	95.98	-
   10000	98.90	98.93	98.65	-
   60000(ALL)	99.64	99.70	99.13	99.32