基于双路径网络的自闭症预测框架

simple, highly modularized framework for autistic prediction is proposed based on dual path network. By stacking several building blocks that aggregate a set of transformations with the same topology, the framework extracts human eye movement information automatically. Residual path and densely connected path are integrated to reuse features at different levels and explore new features. Synthetic oversampling technique is proposed and the form of cost function ischanged to deal with the imbalanced data. By replacing the activation function and adjusting the order of the active layers and the batch normalization layers, the classification accuracy is further improved and the convergence speed is accelerated. The improved dual path network enjoys higher parameter efficiency, lower computational cost and lower memory consumption. Experimental results show that compared with the existing autistic prediction model, the autistic prediction framework based on dual path network has better prediction and optimization performance.