Active One-shot Learning with Memory-Augmented Neural Networks
Abstract
This thesis aims at learning an Active Learning agent for one-shot predictions of texts and images via Reinforcement Learning. There are three different models being implemented, where the baseline is a LSTM network, and the two other are instances of Neural Turing Machines - with different methods for writing to memory. All models are tested against the OMNIGLOT dataset for image classification, and the India News Headlines for text classification. The models are also augmented with a form of margin sampling called Class Margin Sampling, which feature is to select the class of images which constitutes the most difficult episode, from a pool of classes. Furthermore the rewards provided during training is experimented with, to identify how the models alters their behaviour given their learning environment.