Q

[wangxinghangcnn]

Hello! Thank you very much for providing the code. After reproducing your code, I have a question: Could you please tell me where the mapping nodes and enhancement nodes are in the width learning system, and where are they reflected in your code?

[linjiatai]

Hello! Thank you very much for providing the code. After reproducing your code, I have a question: Could you please tell me where the mapping nodes and enhancement nodes are in the width learning system, and where are they reflected in your code?

Thank you for your attention in our work. In fact, this version of PDBL uses the simplified broad learning system of non-enhanced nodes. Here's why:

Broad Learning System is a fast-training classifier which can be divided into three parts include feature mapping, enhancement mapping and output mapping. In enhancement mapping, enhancement nodes based on random weights can be used to increase the feature dimension so as to find the superior hyperplane for classification in higher dimensional space. Since there are enough features extracted by CNN backbone in our method, it is afraid that enhancement nodes cannot bring improvement in the proposed PDBL. To this way, we designed the PDBL without enhanced nodes in this version.

Enhancement mapping is still an important part of broad learning system which not only improve the feature dimension but also equates to randomly weighted combinations of feature nodes. These characteristics may provide some interesting information in some scenarios. For example, we keep appropriate enhancement nodes in our ongoing research on BLS-based weak supervised semantic segmentation task.

I hope my answer will help you. Thanks again for your attention.

[wangxinghangcnn]

Thank you very much for your answer, and also for your excellent work！