Incremental Deep Learning for Object Classification

Kurzfassung

In robotics applications, it is common for tasks to be modified over time. For example, a classification task might be expanded over time to classify more and more classes. The focus of this thesis is to compare the potential of the traditional fine-tuning approach and the newly proposed Progressive Neural Network (PNN) approach [Rusu et al., 2016a] to incrementally adapt a deep predictive model to such dynamically changing tasks. Empirical results are presented in thesis showing that in certain scenarios PNN is significantly more effective than fine-tuning in this regard. To determine these prospective scenarios, where PNN is expected to outperform the fine-tuning approach, this thesis presented a hypothesis which is validated by the experiment results. In addition to that, this thesis also proposed a new method to extend a classifier with new classes with PNN. The experiment results in this thesis showed that it is more effective and reliable than the fine-tuning approach.