Topic description:
Deep neural network have shown a great performance in finding correlation between input variables and a target label. Nowadays, models based on deep learning are usually applied for classification problems. In the case of supervised learning, the model learns from labeled training examples: if the number of samples are large, classification accuracy will rise. However, there are cases when there is a large number of categories, with relatively few samples. Humans are able to learn from only a few examples, an interesting field of research is to create machine learning based solutions to create a similar solution.
The key idea behind one-shot learning is to have a system, where prior knowledge is used as a basis of training, and new categories, new classes, new knowledge is added to this base knowledge during training.
The goal of this thesis is to review current state-of-the-art machine learning techniques for one-shot learning and adjacent fields (e. g. zero-shot learning), design and develop a solution for a one-shot learning task.