Efficient convolutional networks learning through irregular convolutional kernels

As deep neural networks are increasingly used in applications suited for low-power devices, a fundamental dilemma emerges: the trend is to develop models use increasing amount of data, resulting memory-intensive models; however, devices have very limited memory and cannot store large models. Parameters pruning critical model deployment on devices. Existing efforts mainly focus designing highly efficient structures or redundant connections networks. They typically sensitive tasks rely dedicated expensive hashing storage strategies. In this work, we introduce novel approach achieve lightweight from perspective reconstructing structure convolution kernels storage. Our transforms traditional square kernel into line segments, automatically learns proper strategy equipping these segments diverse features. Experimental results show that our can significantly reduce number parameters (pruned 69% DenseNet-40) calculation costs 59% while maintaining acceptable performance (only lose less than 2% accuracy).