Core Concepts
Tensorization of convolution kernels in CNNs allows for effective compression without sacrificing accuracy.
Abstract
The content explores the impact of truncating convolution kernels in dense CNNs, focusing on ResNet-50 models trained on CIFAR-10 and CIFAR-100 datasets. Various tensor network decompositions, such as Tucker and CP, are discussed along with their effectiveness in compressing CNN models. The experiments conducted reveal the robustness of CNNs against correlation truncations based on SVD and CP decompositions. The results show that certain bipartitions allow for significant truncation without a loss in accuracy, indicating the potential for effective compression techniques in neural networks.
I. Introduction
Overview of CNN architecture and image classification tasks.
Evolution of state-of-the-art CNN architectures.
II. Background
A. Dense Convolutional Neural Networks
Description of feature extractor and classifier components.
Explanation of convolution and pooling operations as tensor contractions.
B. Tensor network decompositions of convolution layers
Overview of Tucker, HOSVD, CP decompositions.
Introduction to MPS-based decompositions like tensor train and tensor ring.
III. Truncation of dense CNNs
A. Single-mode truncations
Application of SVD-based truncation to various bipartitions.
B. Two-mode truncations
Similar process as single-mode but with pairs of indices grouped together.
C. MPS-based truncation
D. CP-based truncation
E. Quantifying the impact of truncations
IV. Results for ResNet-50
Setup of the truncation experiments.
Spectra analysis showing flat singular value distributions across different bipartitions.
Stats
No key metrics or figures were provided to support the analysis.