Multi-task Learning Based Classification Algorithm For Balanced Image Data And Unbalanced Temporal Data

In the era of deep learning,inspired by this human learning capability,multitask learning has been proposed in the context of machine learning,aiming to jointly learn multiple related tasks so that the learned information contained in one task can be used by other tasks to improve the generalization ability of all target tasks.In the long-term research process of multitask learning,it is found that the impact of information interaction between tasks is not always positive,and when multiple tasks are not similar,the impact of this information interaction between tasks is a negative effect called negative migration,which is a long-term problem in multitask learning research.At the same time,in the era of deep learning,the overall efficiency of parameter sharing based multitask learning algorithms is low,and it is usually necessary to build a model with a large number of parameters or design a parameter model for each task to realize the parameter sharing mechanism for multitasking.In addition,the balanced and unbalanced characteristics of multi-task data distribution can also have an impact on existing multi-task learning algorithms,and even lead to algorithm failure.In this paper,we will improve the algorithms for multitask learning in three aspects:improving negative migration,enhancing parameter utilization and overcoming algorithm failure in multitask data imbalance scenarios.Our model uses a task-specific binary mask controlled by trainable parameters,i.e.,a mask(mask)for each parameter.It uses the standard backpropagation method to update both the mask and the network weights,reshaping the subnetwork during training.From the perspective of individual network weights,one step of backpropagation not only updates its value,but also changes the task it serves by adjusting the corresponding mask element to 0 or 1.We name this the dynamic fine-grained sharing network(FSN),which implements a dynamic approach for multi-task fine-grained sharing subnetwork is updated in a dynamic manner.In the absence of a priori knowledge about task relatedness,the model can adaptively share more parameters for closely related tasks and fewer parameters for less related tasks,with task-sensitive features.In addition,the model can learn to adjust the sparsity of the subnetwork considering the difficulty of the task.This helps to suppress negative transfer as well as to increase the parameter efficiency to the limit.Due to the task-sensitive nature of the proposed model,this approach is immune to the problem of multi-task learning algorithm failure due to task data imbalance.Experiments on balanced image data show that the dynamic fine-grained shared network model can effectively overcome negative migration and improve parameter utilization,while experiments on unbalanced temporal data show that the method does not fail due to unbalanced multitask distribution data while suppressing negative migration.In addition,this method is applied to the framework of multi-task transfer learning,and Adaptive Transfer Learning via Fine-grained Multi-task Pre-training(ATL)based on fine-grained shared networks is proposed.The method uses the proposed dynamic fine-grained shared network(FSN)with high sensitivity to the task in the pre-training phase,and trains a subnet mask for each task in addition to the network weights,and selects the most appropriate subnet for each downstream task in the fine-tuning phase.Thus,different downstream tasks can be fine-tuned according to different network structures and benefit from the results of their most closely related pre-trained tasks.According to our experiments,APT outperforms the traditional pre-training baseline.Moreover,experiments show that the quality of network weights pre-trained by FSN alone is much higher even without sub-network adaptation and can be used within the framework of traditional pre-training to improve performance.