End-to-end data-dependent routing in multi-path neural networks

Neural networks are known to give better performance with increased depth due their ability learn more abstract features. Although the deepening of has been well established, there is still room for efficient feature extraction within a layer, which would reduce need mere parameter increment. The conventional widening by having filters in each layer introduces quadratic increment parameters. Having multiple parallel convolutional/dense operations solves this problem, but without any context-dependent allocation input among these operations: computations tend similar features making process less effective. Therefore, we propose use multi-path neural data-dependent resource from layers, also lets an be routed end-to-end through paths. To do this, first introduce cross-prediction-based algorithm between tensors subsequent layers. Second, further routing overhead introducing feature-dependent cross-connections successive Using image recognition tasks, show that our superior existing and adaptive extraction, even ensembles deeper at complexity.