Broad Learning for Robotic Material Recognition

Abstract

Broad learning was proposed as a flat network based on a random vector functional link neural network. The horizontal expansion of the network is realized by the increment of the feature nodes and enhancement nodes, which shows a good learning ability. In addition, fast incremental learning algorithms are implemented in horizontal expansion without the complete retraining of the entire network when structure is not enough to model the system. However, due to the design of its single layer neural network and the linear mapping to form feature nodes, it is difficult to extract high-level abstract features if inputs are more complicated. To break through this limitation, two algorithms C-CFBRL and C-LCFBRL are proposed based on CFBRL and LCFBRL, combining with convolution and pooling to automatically extract the key features of the image. The related experiments are conducted on the datasets which include RGB images, Depth images, and contact microphone signals. It is proved that the proposed BRL models and algorithms are very effective in classification. It is also shown that broad learning can be easily extended to other networks to solve related problems in robotics.