Few-Shot and Transfer Learning with Manifold Distributed Datasets

A manifold distributed dataset with limited labels makes it difficult to train a high-mean accuracy classifier. Transfer learning is beneficial in such circumstances. For transfer learning to succeed, the target and base datasets should have a similar manifold structure. A novel method is presented in this paper for determining the similarity between two manifold structures. To determine whether target and base datasets have similar manifolds and are suitable for transfer learning, this method can be used. A novel few-shot algorithm is then presented that uses transfer learning to classify manifold distributed datasets with a limited number of labels. Using the base and target datasets, the manifold structure and its relevant label distribution are learned. Using this information in combination with the few labels and unlabeled data from the target dataset, we can develop a classifier with high mean accuracy.