Incremental Learning in Hierarchical Neural Networks for Object Recognition

Rebecca Fay, Friedhelm Schwenker, Günther Palm

Robots that perform non-trivial tasks in real-world environments are likely to encounter objects they have not seen before. Thus the ability to learn new objects is an essential skill for advanced mobile service robots. The model presented in this paper has the ability to learn new objects it is shown during run time. This improves the adaptability of the approach and thus enables the robot to adjust to new situations. The intention is to verify whether and how well hierarchical neural networks are suited for this purpose. The experiments conducted to answer this question showed that the proposed incremental learning approach is applicable for hierarchical neural networks and provides satisfactory classification results.