LA-ReduNet Achieves Lightweight Adaptive Neural Network Architecture.

ReduNet, a white-box neural network, uses the Maximal Coding Rate Reduction (MCR^2) principle to create low-dimensional, discriminative features through a layer-wise construction process. Unlike traditional deep learning models that rely on backpropagation, ReduNet explicitly derives layer parameters from preceding layers, offering a transparent and mathematically interpretable approach. However, this layer-wise construction often demands a large number of layers for the MCR^2 objective to stabilize, leading to substantial parameter storage requirements. To overcome this limitation, researchers have developed LA-ReduNet, a lightweight and adaptive architecture. LA-ReduNet refines the original layer-wise update rule by incorporating hyperspherical manifold learning and adaptive step sizes. This innovation allows the model to achieve stable discriminative feature representations with significantly fewer unfolded layers. Experimental results demonstrate that LA-ReduNet can reduce the number of required layers by an order of magnitude, drastically cutting down parameter storage—by approximately 1/29th in some settings—while preserving classification accuracy comparable to the original ReduNet. This advancement makes ReduNet-based models more efficient and practical for deployment.