Learn from one image: Dynamic One-shot learning based on parameter generation

Abstract

State-of-the-art deep learning algorithms are usually pre-trained on datasets containing millions of images. Adding new classes to these pre-trained networks, require large number of images for each of the new classes. Formulation of such large scale datasets usually require a lot of effort and time. The aim of this paper is to develop novel deep learning based one-shot learning framework which can achieve state-of-the-art results on new classes (one-shot classes) which have only one image each during the training phase. Adding these new one-shot classes, should not degrade the performance of the model on pre-trained classes. Multi-layer transformation function has been proposed in this paper for one-shot learning, where activations of a class are converted to their corresponding parameters. The model is pre-trained on large-scale base classes and the model adapts to new classes with zero training. Experiments were conducted on opensource datasets like MiniImageNet and Pascal-VOC using Nvidia K80 GPU. The model achieves an accuracy of 93.14% for large scale base classes and 64.69% for one-shot classes which is more than 3% better than the current state-of-the-art models.