Universal structural patterns in sparse recurrent neural networks

Sparse neural networks can achieve performance comparable to fully connected networks but need less energy and memory, showing great promise for deploying artificial intelligence in resource-limited devices. While significant progress has been made in recent years in developing approaches to sparsify neural networks, artificial neural networks are notorious as black boxes, and it remains an open question whether well-performing neural networks have common structural features. Here, we analyze the evolution of recurrent neural networks (RNNs) trained by different sparsification strategies and for different tasks, and explore the topological regularities of these sparsified networks. We find that the optimized sparse topologies share a universal pattern of signed motifs, RNNs evolve towards structurally balanced configurations during sparsification, and structural balance can improve the performance of sparse RNNs in a variety of tasks. Such structural balance patterns also emerge in other state-of-the-art models, including neural ordinary differential equation networks and continuous-time RNNs. Taken together, our findings not only reveal universal structural features accompanying optimized network sparsification but also offer an avenue for optimal architecture searching.

Xin-Jie Zhang, Jack Murdoch Moore, Gang Yan, Xiang Li