Modulated Convolutional Networks

While the deep convolutional neural network (DCNN) has achieved overwhelming success in various vision tasks, its heavy computational and storage overhead hinders practical use of resource-constrained devices. Recently, compressing DCNN models attracted increasing attention, where binarization-based schemes have generated great research popularity due to their high compression rate. In this article, we propose modulated networks (MCNs) obtain binarized DCNNs with performance. We lead a new architecture MCNs efficiently fuse multiple features achieve similar performance as full-precision model. The calculation is theoretically reformulated discrete optimization problem build DCNNs, for first time, which jointly consider filter loss, center softmax loss unified framework. Our are generic can decompose filters e.g., conventional VGG, AlexNet, ResNets, or Wide-ResNets, into compact set optimized based on projection function updated rule during backpropagation. Moreover, modulation (M-Filters) recover from ones, specific calculate proposed substantially reduce cost by factor 32 comparable counterparts, achieving much better than other state-of-the-art models.