In this project, I have used three different contrastive learning algorithms to produce an informative embedding of the Fashion MNIST dataset. Each algorithm uses an open-source augmentation module from the PyTorch library and a simple convolutional neural network embedding scheme. The first two algorithms are applied in the self-supervised setting, and the third algorithm is applied in the supervised setting. I measure the quality of each embedding through a series of downstream classification and clustering tasks.
The results of the first algorithm, triplet loss, is shown in the self_triplet.ipynb notebook. The results of the second
algorithm, self-supervised contrastive loss, is shown in the self_supcon.ipynb notebook. The results of the final algorithm,
supervised contrastive loss, is shown in the supcon.ipynb notebook. The first algorithm is specified in
FaceNet: A Unified Embedding for Face Recognition and Clustering by Florian Schroff, Dmitry Kalenichenko, and James Philbin.
The other two algorithms are specified in Supervised Contrastive Learning by Khosla et al.
A detailed written report specifying each of my experiments can be found within the Report directory of this repo. The Report
directory also includes a brief 10 minute presentation of my motivation and conclusions from this project.