基于张量结构的深度卷积神经网络轻量化模型设计

Deep convolutional neural networks have achieved state-of-the-art results on many tasks including image and video processing and recognition. However, the complexity of the network structure also brings large computational amount, and blocks to the research and practical applications of deep learning. How to improve the computing efficiency but not to reduce the network performance is the key issue of lightweight model design. Aiming at the problem of redundant information in weight tensor and insufficient information integration of group convolution, this project uses tensor structure characteristics of weights and features to study the design of lightweight network. Specifically, the project's research includes: 1) Designing a lightweight alternative to convolution by using "low-rank + sparse" tensor decomposition to more accurately reduce redundancy within the convolutional layer; 2) In order to reduce the redundancy between convolutional layers, we use "low-rank + sparse" tensor decomposition for efficient weight usage by sharing the bases of tensor subspace, and simultaneously keep the specific information within layers; 3) In order to solve the problem of insufficient information fusion between groups in group convolution, a high-order tensor is constructed to apply high-order information fusion group convolution. The research of this project will provide theoretical support for the design of lightweight convolutional neural network structure, speed up the network training and testing process, reduce the threshold of technical usage, and promote the real-time application of deep learning on user's small terminal device.

深度卷积神经网络在图像、视频的处理与识别等许多任务上都取得当前最好结果。但其复杂的网络结构也带来巨大计算量，给深度学习的研究与实际应用带来挑战。如何在提升计算效率时尽量不降低网络性能是轻量化模型设计的关键问题。针对权重张量的冗余信息和分组卷积信息融合不充分的问题，本项目利用权重和特征的张量结构特点，研究轻量化卷积设计网络的设计。具体来说，本项目的研究包括：1）利用“低秩+稀疏”的张量分解形式，更准确地减少卷积层内的冗余，设计卷积的轻量化替代结构；2）针对卷积层间的冗余，利用“低秩+稀疏”的张量分解形式，通过共享张量子空间的基来实现参数共享的,同时保留了层内的特异性信息；3）针对分组卷积组间信息融合不充分的问题，构建高阶张量来实现高阶信息融合分组卷积。本项目的研究将为轻量化卷积神经网络结构的设计提供理论支持，加快网络的训练与测试过程，降低技术使用门槛，促进深度学习在用户小型终端的实时应用。

卷积神经网络在各类人工智能应用中占据重要地位，但大量的计算量耗时耗能，阻碍了其在小型终端的应用。本项目针对卷积层内冗余提出层内局部稀疏、低秩与稀疏分解等轻量结构设计方法；针对层间冗余，提出全局稀疏和权重二阶导的重要性度量以及网络压缩超参的自适应搜索学习等轻量结构设计方法；针对分组的高阶权重张量，提出分组二值化网络、基于注意力机制的动态通道重排、基于可学习位移的通道重排与融合等轻量结构设计方法；将轻量模型研究过程中所用到的张量结构、低秩分解、高阶融合等思想用于脑电信号情感计算、神经网络训练方法设计、人脸表情识别等拓展应用。本项目共发表SCI或EI检索论文9篇、申请发明专利9项，其中JCR一区期刊论文3篇、CCF A类会议论文1篇、获得授权发明专利1项。本项目针对卷积神经网络的冗余，设计轻量模型并促进轻量模型信息融合，有效提升轻量模型的性能，为卷积神经网络在小型终端的广泛应用以及“双碳”目标的达成做出贡献。