Development of Conceptual Learning Model Based on Various Stability Features

To promote efficient and flexible learning, a conceptual learning model can be developed using the differences between the stable features of various concepts, thereby allowing scalable and continuous learning concepts to be obtained from a small number of labeled samples. By using a stable feature expressed as a vector to represent each concept, the conceptual learning model can distinguish the representative feature of each concept, thus increasing the interpretability of the model. In addition to adjusting the mapping relationship between concept instances and concept stability features, constraints on the differences between stable features of different concepts are also introduced into the model. Because this constraint can improve the sensitivity of the instance features to the stable features of each concept, the proposed model has the ability to learn from a small number of samples. In this paper, sub-networks of the same structure were applied to learn each concept, and unified learning methods were used to achieve scalability and the ability to easily manage new concepts. By training the proposed model using the MNIST dataset, the model demonstrates that prediction accuracy of 96% can be achieved as a result of using 100 labeled training samples for each concept. Furthermore, the proposed model not only demonstrated an improved learning ability for small samples, but it was also found to considerably reduce the training time, with each concept requiring only 18 s on average. Thus, the small-sample and continuous learning capabilities of the proposed model enable the construction of a conceptual space that promotes improved knowledge representation and complex scene understanding.