shreyas-bhat / stl10

Comparing different learning paradigms on the STL 10 dataset and carrying further analysis in each method

STL10

• Performed different learning procedures on the STL10 dataset - supervised learning, semi-supervised learning and self-supervised learning

Supervised Learning

• Used ResNet-50 Architecture and got validation accuracy of 68.7

Semi-Supervised Learning

• Used Pseudo-Labeling method using the same encoder architecture as in supervised learning

Model	Supervised Validation Accuracy	Semi-Supervised Validation Accuracy	Change in Accuracy
CNN Model	59.4	64.62	5.08
ResNet-50 Model	68.73	72	3.27

Self-Supervised Learning

• For this I used the SimClr framework for contrastive learning and get a valiation accuracy of 53.30%

AutoAugment

• I tried to implement semi-supervised tasks using SimClr and augment images using AutoAugment method. The operations we will be using are shearing, translating, rotation, auto_contrasting, brightness, sharpness, cutout, etc., and the policies for each augmentation are selected randomly and applied in our dataset for producing image augmentations