Tensor decomposition to compress convolutional layers in deep learning

Feature extraction for tensor data serves as an important step in many tasks such anomaly detection, process monitoring, image classification, and quality control. Although methods have been proposed feature extraction, there are still two challenges that need to be addressed: (i) how reduce the computation cost high dimensional large volume data; (ii) interpret output features evaluate their significance. The most recent deep learning, Convolutional Neural Network, shown outstanding performance analyzing data, but wide adoption is hindered by model complexity lack of interpretability. To fill this research gap, we propose use CP-decomposition approximately compress convolutional layer (CPAC-Conv layer) learning. contributions our work include three aspects: adapt kernels derive expressions forward backward propagations CPAC-Conv layer; compared with original layer, can number parameters without decaying prediction performance. It combine other layers build novel Deep Networks; (iii) value decomposed indicates significance corresponding map, which provides us insights guide selection.