A use-case of semi-supervised learning for images

[sayakpaul]

@arjung since semi-supervision has come a long way in computer vision. Since Neural Structured Learning has first-class support to perform semi-supervision I was thinking of example doing the following:

• Take the 10000 images randomly from CIFAR10 and also split the rest of the samples into unlabeled and validation samples.
• Train a model on this subset. This model will be some pre-trained ImageNet-1k model fine-tuned on the CIFAR10 10000 samples I mentioned.
• Use this model (without the final classification layer) as a feature extractor to generate embeddings for the subset and rest of the 40000 images in CIFAR10.
• Build the graph augmented with structural similarity.
• Train another model with this graph data.
• Evaluate the model on the test set.

What do you think of this?