Sie befinden Sich nicht im Netzwerk der Universität Paderborn. Der Zugriff auf elektronische Ressourcen ist gegebenenfalls nur via VPN oder Shibboleth (DFN-AAI) möglich. mehr Informationen...
Though pointwise convolutions play critical roles in compact networks such as ShuffleNet [1] and MobileNet Series [2], [3], unfortunately, they typically account for the majority of the parameters of the entire network, preventing further reductions in model size. To address this shortcoming while retaining model performance, we propose a brand new pointwise convolutional block, termed as the "Improved Pointwise Convolution" (IPC) block. For each channel feature of the input tensor in IPC block, the corresponding scores are first adaptively computed and then all of the scores are used to perform pointwise convolution with the input tensor. We design an effective mechanism to construct the IPC block, which therefore provide three significant advantages: (i) The parameters of the regular pointwise convolu-tions no longer need to be stored on physical devices, thus contributing to remarkably reduction in model size. (ii) The tradeoff between model size and performance can be controlled by adjusting two hyper-parameters in the IPC block. (iii) The IPC block can be a substitution for regular pointwise convolution in neural networks especially in compact networks. Compared to other popular compact models, experiments on CFAR-10 demonstrate that the IPC block has the superior ability in further model compression without affecting their accuracy.