Research On Lightweight Multidimensional Convolutional Neural Networks

Two dimensional Convolutional Neural Networks(Conv Nets)have been widely adopted as powerful models and have achieved state-of-the-art performance in many image related tasks.However,their extensions,e.g.,three dimensional Conv Nets,are still struggling for leading performance for high dimensional(HD)signal processing,partially due to the explosion of training parameters,greatly enhanced computational complexity and memory cost.In this paper,we present two kinds of simple,lightweight,yet efficient Conv Nets for the HD signal processing by allowing a separable structure on the Conv Net throughout the learning process.First we propose multi-dimensional separable Conv Nets with sequential structure,in which a series of alternate one dimensional convolution operations take place of conventional ND convolution operation.This makes it not only be compatible with multi-dimensional signals,but also decrease the computational complexity of networks from exponential to linear in contrast with traditional Conv Nets.Multiple experiments between the proposed lightweight Conv Net and the traditional one on both MNIST dataset and KTH dataset are implemented.Experimental results indicate that the proposed Conv Net is competitive with the traditional Conv Net in recognition accuracy,but more robust against overfitting.Second we propose multi-dimensional separable Conv Nets with multi-stream parallel structure.The proposed model utilizes high-order tensors to represent multidimensional signals,and consists of multi-level modules whose inputs and outputs are high-order tensors.In each module,the characteristics of the signal in different dimensions are extracted by tensor operation and a set of one-dimensional convolutional parallel operations.Through feature fusion,the multi-dimensional signal characteristics of the module are obtained.It can not only extract the hierarchical characteristics of multi-dimensional signals from low-level to high-level,but also greatly reduce the number of network parameters,reduce the space-time complexity of network training,and improve the inference speed of network.In addition,a series of one-dimensional convolution operations are utilized to increase the number of training samples in each dimension,reducing the sampling complexity.The proposed model has been evaluated on both 2D and 3D benchmarks CIFAR-10,MNIST and KTH.Experimental results show that the presented model achieves competitive performance with greatly reduced computational and memory costs in comparison with the state-ofthe-art Conv Net models.On CIFAR-10,the proposed model at a size of 0.47 M achieves high score in accuracy while VGG network achieves about a same score in accuracy at a size of 138 M.We believe the proposed model provides an alternative and promising way to handle multi-dimensional signals especially for low-end devices.