Knowledge-based hybrid connectionist models for morphologic reasoning

Texture morphology perception is essential feedback for robots in tactile-related tasks (such as robot's electrical palpation, manipulation, or recognition of objects in complex, wet, and dark work conditions). However, it is tough to quantify morphologic information and define morphologic feature. For this reason, it is difficult to use prior tactile experience in detection, which results in large dataset requirements, time costs, and frequent model retraining for new targets. This study introduced a hybrid connectionist symbolic model (HCSM) that integrates prior symbolic human experience and the end-to-end neural network. HCSM requires smaller datasets owing to using a symbolic model based on human knowledge. Moreover, HCSM improves the transferability of detection and interpretation of recognition results. The neural network has the advantage of easy training. The HCSM combines the merits of both connectionist and symbolic models. We have implemented tactile morphologic detection of basic geometry textures (such as bulges and ridges) using the HCSM method. The trained model can be transferred to detect gaps and holes by manual adjustment of the symbolic definition, without model retraining. Similarly, other new morphology can be detected by only modifying the symbolic model. We have compared the recognition performance of the proposed model with that of the traditional classification models, such as LeNet, VGG16, ResNet, XGBoost, and DenseNet. The proposed HCSM model has achieved the best recognition accuracy. Besides, compared with classic classification models, our method is less likely to misrecognize one target as a completely different counterpart, providing a guarantee for generalization boundaries of recognition to a certain degree.