Image Classification Based On Incremental Learning Method Of Task Sorting

Artificial intelligence technology is a major driver of the new technological revolution and industrial transformation,with a role and depth comparable to that of previous industrial revolutions.The connection between brain science and artificial intelligence can be clearly seen in the history of AI development,and many of the pioneering AI scientists are also brain scientists.The purpose of AI research is to conduct theoretical research and develop computer systems that can perform tasks in place of human intelligence and that have functions such as perception,recognition,decision making,and control.The next generation of AI should be capable of incremental learning,like human intelligence.It can learn a new task while maintaining memory of the learned task,thus enabling incremental learning of multiple tasks.When humans learn knowledge with associated rules,they improve learning by finding the best learning sequence.Inspired by the fact that the brain is reorganized in an orderly manner during learning,this paper exploring the effect of task order on learning outcomes in incremental learning by artificial neural networks.We find that in incremental learning,different task order produces different learning results,and the task order of learning is very important.In order to deeply investigate the effect of different task order on incremental learning results,the In this paper,we design a task ordering module and propose an incremental learning method based on task sorting.The method requires designing a task sorting module first and then implementing it by embedding the module into the incremental learning method.The task sorting module uses the features of the tasks to achieve the sorting of the tasks.In this paper,the task sorting module is designed into three parts,which are feature extraction of images,similarity calculation between features and task sorting algorithm.Among them,the The feature extraction of image is realized by convolutional neural network,a two-dimensional image will get a vector data,i.e.features,after convolutional neural network.In order to better reflect the difference of features in different directions,this paper uses cosine similarity to calculate the similarity between features.In the task sorting algorithm,this paper searches for the desired task order by setting different similarity relationships between tasks.In order that the final search results converge to the global optimal results and avoid being limited to local minima in the search process,the simulated annealing algorithm is used to achieve the task order search.At the same time,inspired by the fact that humans learn multiple tasks in succession by controlling the complexity of each task to improve the learning effect,this paper explores the need for artificial neural networks to be more efficient and more effective.In this paper,we explore the effect of the number of tasks to be learned incrementally in artificial neural networks on the incremental learning results.Finally,in order to verify the effectiveness of the incremental learning method based on task ranking,three task similarities and three task numbers are set in this paper,and experiments are conducted on the image classification task.The experimental data show that that different task sequences produce different incremental learning results.The proposed task sorting module can search for the desired task order based on the similarity relationship between features.There is a positive correlation between the similarity relationship between task order and the accuracy of incremental learning.That is,incremental learning by task order with high similarity yields the best learning results.The largest difference in average accuracy is 4.59%compared with that of low similarity.In addition,the experimental data also show that the artificial neural network is similar to human learning in terms of the number of categories that need to be learned for a single task,and the smaller the number of image categories that need to be learned for a single task,the better the accuracy of incremental learning,with the largest difference in the average accuracy of 9.59%.The above results indicate that artificial neural networks have some similarity with human brain in incremental learning,and incremental learning by higher similarity or less number of tasks will get better learning results,and the idea of studying incremental learning by the similarity relationship between tasks and the number of tasks is valid in this paper.