يادگيري چندوظيفه‌اي هندسي براي دسته‌بندي با مجموعه‌داده‌ي محدود

Despite the ever-increasing expansion of learning models and the great thirst for collecting massive data to train more complex and comprehensive models, today the development and design of learning models using a limited amount of data is still an active and prominent research field in artificial intelligence research. Lim-ited dataset learning solutions are of great importance in many real-world prob-lems where collecting large amounts of data is impossible, costly, or time-consuming. In this research, an attempt has been made to provide a solution to im-prove learning models with a limited data set. Limited data learning models are usually presented in the form of multi-task learning. By using the multi-task learn-ing approach, a model can be trained to perform several tasks defined in a com-mon feature space and maximize the efficiency of the data. It is also possible to solve the problem of lack of data to some extent. The main idea of the multitask learning model is that by sharing information between tasks, it can perform the learning process more efficiently and effectively than when each task is taught separately. In this research, a multi-task learning framework is proposed for classi-fication with a limited dataset. Furthermore, in many real-world applied problems, the tasks may be structured in such a way that they can be solved by a geometric approach. In the field of machine vision, issues such as object recognition, facial rotation estimation, facial expression recognition, and even facial expression re-production all involve reasoning about the geometric structure of images; There-fore, in the conducted research, the geometric features of the data are also taken into consideration and a geometric multi-task learning approach is presented. In geometric multi-task learning, the data related to different tasks can be defined in a regular Euclidean space or a non-Euclidean space such as a manifold or specifi-cally a graph. In this research, the most important tool for analyzing and modeling data with a graph structure, namely graph neural networks, has been used. Despite the need for these networks for a lot of data to learn the representation, it has been tried to minimize this dependence by using a multi-task learning approach and to increase the efficiency of the data by using the existing geometric features be-tween different tasks. In the proposed approach, we learn data representation by using a self-encoding graph network and use the obtained representations in the hidden data space in the classification problem with a limited data set. By eva‎luat-ing the results of the proposed framework, it is possible to see a significant im-provement and superiority among multitasking learning approaches.

6

33583