Supervised, Self-supervised and Semi-Supervised Learning

Under Prof. Tirtharaj Dash

As part of the course project work for CS F425 Deep Learning, I worked on examining several learning modalities on the STL-10 Dataset.

As part of this group project, I contributed to "Self-Supervised and Semi-Supervised Learning on STL-10 Dataset." To train a supervised model (ResNet9) with a 74 % accuracy, we applied different data-augmentation methodologies to enhance the labeled data of 5,000 photographs. Then, using the self-supervised Barlow Twins technique, we generated approximation embeddings and trained a linear classifier with a threshold to obtain 76 % accuracy on unlabeled data. By pseudo-labelling the data using the labeled data on the ResNet9 architecture with a threshold, we used semi-supervised learning to train
ResNet12 to achieve a 73 % accuracy on testing data. I learned how to use PyTorch effectively, understood
the differences between different learning methods, and discovered that cross-entropy loss has a softmax
function built-in :) https://drive.google.com/file/d/1jF1s2GEvCZ5NpVR0oT1Z0HeWDSECM9pX/view?usp=sharing