Compression of Convolutional Neural Networks Employing Tensor Train and High Dimensional Model Representation

Deploying deep neural networks on edge devices requires compact and efficient models. This study investigates the compression of convolutional neural networks using Tensor Train (TT) and High Dimensional Model Representation (HDMR). TT reduces parameter count by factorizing weight tensors into lowrank cores, while HDMR filters high-order interactions to enhance interpretability. We evaluate four strategies - TT, HDMR, and TT HDMR - on both ResNet and VGG architectures using the CIFAR-10 dataset. Experimental results show that TT HDMR achieves the best compression-accuracy balance, offering significant size reduction with minimal performance drop. This hybrid method can potentially combine structural and functional decompositions for efficient and interpretable deep learning.