A Mathematical Theory of Reusable Neural Bases for Network Compression
The paper proposes compressing networks by rebuilding each block from shared neural “bases.”
Binshuai Wang introduces LRNBA, a linear reusable neural bases architecture meant to cut parameter and memory costs in training and inference. The setup represents network blocks as linear combinations of a common basis set, drawing inspiration from RNN-style reuse. The abstract says this lets models become wider and deeper under the same parameter budget. Reported experiments show stable training, comparable or faster convergence, and lower loss than classical architectures. ArXiv · AI/CL/LG's note
Binshuai Wang introduces LRNBA, a linear reusable neural bases architecture meant to cut parameter and memory costs in training and inference. The setup represents network blocks as linear combinations of a common basis set, drawing inspiration from RNN-style reuse. The abstract says this lets models become wider and deeper under the same parameter budget. Reported experiments show stable training, comparable or faster convergence, and lower loss than classical architectures. ArXiv · AI/CL/LG's note
score 4