关键词: 渐进式神经网络, 深度神经网络, 持续学习, 相关性差异, 复杂多任务

Abstract: Classical Progressive Neural Network(PNN) improves the performance of neural networks on the current task by acquiring empirical knowledge of previous tasks,but ignores the influence of the correlation difference between progressive tasks on the performance of the network when there are many progressive tasks. For such a scenario with a large number of progressive tasks and differences in the correlation between tasks, a Progressive Neural Network based on Correlation Differentiation Transfer(CDT-PNN) algorithm was proposed. Firstly,the correlation of the two progressive tasks was first evaluated using a cosine similarity-based algorithm.Then,the knowledge parameter transfer of the neural network was determined by exploiting the correlation between the current task and the previous task.The previous asymptotics that were negatively correlated with the current progressive task were removed.Finally,the knowledge parameters of the tasks were randomly selected according to the correlation between tasks and the knowledge parameters of the progressive tasks were randomly selected to transfer the parameters. Experiments were conducted on the cifar-100 dataset and mnist handwriting recognition dataset with different levels of noise.The experimental results show that CDT-PNN performs better than PNN in complex multi-task scenarios. The average classification accuracy of the experimental tasks on cifar-100 dataset is increased by 6.6 percentage points, and that on mnist dataset is increased by 1.56 percentage points.

Key words: progressive neural network, deep neural network, continual learning, correlation differentiation, complex multi-task