Update #4

08 Oct 2018

Done

• Experiments with USPS dataset
• Scoped out possible datasets for future (FastText and 20 newsgroups)

Objective:

Can a graph net trained on MNIST data generalize well to the US Postal Service (USPS) dataset, a similar dataset involving handwritten digits?
Size of training set: 7,291 images
Size of test set: 2,007 images

Experiment 1: Pre-trained MNIST net

Can the graph net pre-trained on MNIST also generate good embeddings on USPS? The initial results are not very positive, but we observe a small amount of clustering.

Experiment 2: Fine-tuning the pre-trained network on USPS data

Starting from the pre-trained network, we fine-tune the weights of the network by re-training on images from the USPS dataset.