Approach:
We use dataset-distillation to generate 10 MNIST one-shot image samples, and save the image-labels in this pickle file. We then use Tensorflow-Keras to train a model using the synthetic images, and evaluate the performace using the MNIST testing dataset.
Training for 50 epochs with a batch size of 10, we get an average of around 34 ± 13 % accuracy (avg of 50 runs) on the MNIST testing dataset.