In the Jupyter notebook for this repository, I begin by calculating the bottleneck features for the CIFAR-10 dataset.
These features are then visualized with a Barnes-Hut implementation of t-SNE, which is the fastest t-SNE implementation to date.
For Python >= 3.5, this implementation can be installed by running
pip install git+https://github.com/alexisbcook/tsne.git
The bottleneck features are then fed to a shallow CNN for image classification. The model attains a test accuracy of 82.68%.
This repository is discussed in the blog post here.