Few-shot learning of deep convolutional models

Frano Rajič

Report

Image classification is a complex problem with many interesting applications. Convolutional neural networks have achieved superior results in many areas of computer vision, especially in image classification tasks. However, classical approaches to learning convolutional models require huge amounts of labeled data for each semantic class. We relaxed this requirement with a deep convolutional Siamese neural network that supports learning from a small number of examples. The proposed model is based on deep correspondence embeddings that map specimens of the same semantic classes into a compact region of multiplicity of the most abstract latent representation. We evaluate the model on the Omniglot dataset.