Convolutional Neural Networks with analytically determined Filters
Matthias Kissel, Klaus J. Diepold · 2022 International Joint Conference on Neural Networks (IJCNN) · 2022
In this paper, we propose a new training algorithm for Convolutional Neural Networks (CNNs) based on well-known training methods for neural networks with random weights. Our algorithm analytically determines the filters of the convolutional layers by solving a least squares problem using the Moore-Penrose generalized inverse. The resulting algorithm does not suffer from convergence issues and the training time is drastically reduced compared to traditional CNN training using gradient descent. We validate our algorithm with several standard datasets (MNIST, FashionMNIST and CIFAR10) and show that CNNs trained with our method outperform previous approaches with random or unsupervisedly learned filters in terms of test pre-diction accuracy. Moreover, our approach is up to 25 times faster than training CNNs with equivalent architecture using a gradient-descent based algorithm.