Update #6

05 Nov 2018

Experiment 1: Visualising ImageNet embeddings

A pre-trained ResNet was used to obtain 512-dimensional feature vectors of the following ImageNet samples.

Data Summary

Training Data

Test Data

A graph net was trained on the t-SNE embeddings of the training data. Here are the visualisations produced:

Results Summary

• t-SNE performs much better than the graph net, but it is more computationally expensive.
• The performance of PCA seems comparable to that of the graph net, but it does not achieve good separation on imbalanced class distribution.