One of the main obstacles to the spread of deep learning today is its restriction to closed world scenarios. However, several solutions have been proposed to address the problem: in this paper, we analyse some state-of-the-art methods capable of operating in an open world in image classification. We apply slight variations to the best known methods and then propose some of our own modifications. This work is carried out along two main axes: * incremental learning: to gradually learn new classes; * open set: to recognise whether the classes belong to the knowledge that the model already possesses.
Referring to the paper, the repository is organised as follows:
- baseline contains code for Finetuning, LwF and iCaRL, referring to Sections 4.1 and 4.2
- ablation_study refers to Section 4.3
- open_world refers to Section 4.4
- our_modifications refers to Section 4.5