Research On Key Techniques Of Multi-task Lifelong Learning Based On Knowledge Replay

With the rapid development of artificial intelligence(AI)technology,machine learning has made significant progress in fields such as image classification,object detection,and image segmentation.However,in real life,the data we obtain is often streaming,and data for different tasks may arrive sequentially over time,making it necessary for machine learning algorithms to have the ability of lifelong learning.Lifelong learning,also known as continual learning or incremental learning,enables machine learning algorithms to continuously learn new features from new data while also retaining the features learned from previously acquired data.This is currently one of the hottest research directions in the field of machine learning.Currently,many researchers have begun to study lifelong learning in machine learning and have proposed many methods that can effectively suppress catastrophic forgetting,which have been applied in image classification,semantic segmentation,and object detection.Current lifelong learning mainly focuses on homogeneous multi-task scenarios,but in reality,there are often coexisting scenarios of different computer vision tasks,so how to maintain lifelong learning in heterogeneous multi-task scenarios is a major challenge in current research.This paper focuses on lifelong learning methods in heterogeneous multi-task scenarios.First,this paper studies and implements a lifelong learning framework based on heterogeneous multi-task scenarios.Based on three computer vision task scenarios of image classification,semantic segmentation and target detection,this paper proposes a lifelong learning heterogeneous multi-task scenario,and builds a convolutional neural network HCNN that can complete heterogeneous multi-task learning.This paper proves through experiments that the HCNN network can complete lifelong learning in these three scenarios,and catastrophic forgetting will occur in the heterogeneous multi-task lifelong learning scenario,which lays the foundation for subsequent research on heterogeneous multi-task lifelong learning.Then,based on the established heterogeneous multi-task lifelong learning framework,this paper researches and implements the lifelong learning method HSR-LL based on knowledge replay.The HSR-LL method proposes a more applicable representative sample storage strategy in heterogeneous multi-task scenarios,and optimizes the knowledge distillation loss to establish a heterogeneous multi-task scenario loss.This paper also verifies the advanced nature of the knowledge replay-based method and the applicability of the representative sample storage strategy selected by the HSR-LL method through experiments.Subsequently,this paper also further experimentally verified the values of the key parameters"sample storage space size" and "knowledge distillation loss weight".This paper selects several classic methods in the field of lifelong learning to compare with the HSR-LL method,and the experiment proves the advanced nature of the HSR-LL method.Finally,based on the HSR-LL method,this paper studies and implements Attention-PLL,an attention-based lifelong learning method for progressive heterogeneous tasks.The Attention-PLL method uses the attention mechanism to select the key nodes of each task in the network layer,builds differentiated network nodes for each task,and further reduces the mutual influence between tasks.The experimental results show that the Attention-PLL method further improves the performance.This paper also provides more ideas for further research on lifelong learning in heterogeneous multi-task scenarios.