Omniglot, the “transpose” of MNIST, with 1623 character classes, each with 20 examples.
Use background set of 30 alphabets for training and evaluate on set of 20 alphabets. Refer to this script for sampling setup.
Report one-shot classification (20-way) results using a meta learning approach like MAML.
Apply an automatic portrait segmentation model (aka image matting) to celebrity face dataset.
Food-101 is a challenging vision problem, but everyone can relate to it. Recent SoTA is ~80% top-1, 90% top-5. These approaches rely on lots of TTA, large networks and even novel architectures.
Train a decent model >85% accuracy for top-1 for the test set, using a ResNet50 or smaller network with a reasonable set of augmentations.
Apply a supervised or semi-supervised ULMFiT model to Twitter US Airlines Sentiment.