Two-Valued Symmetric Circulant Matrices for Efficient Deep Learning
A recent study published on arXiv (2605.16443) presents the Two-Valued Symmetric Circulant Matrix (TVSCM), an innovative sparse framework for deep neural networks that employs merely two weights per layer to preserve both circulant and symmetric characteristics. This novel design significantly lessens the need for storage and simplifies computational processes, thereby overcoming challenges faced by resource-constrained platforms, including edge devices. In contrast to conventional sparse learning methods like pruning and low-rank approximation, TVSCM offers a highly structured form of sparsity without necessitating extra hardware or additional stages. The focus is on fully connected layers, which usually demand numerous weights, posing difficulties for edge devices. The implications of this research extend to vision, medical diagnosis, and IoT applications.
Key facts
- Paper arXiv:2605.16443 proposes Two-Valued Symmetric Circulant Matrix (TVSCM).
- TVSCM uses only two weights per layer for circulant and symmetric structure.
- Architecture targets fully connected layers in deep neural networks.
- Aims to reduce storage and computational complexity for edge devices.
- Avoids traditional methods like low-rank approximation and pruning.
- Applications include vision, medical diagnosis, and IoT.
- TVSCM provides extreme structured sparsity without extra hardware.
- Paper addresses high storage requirements of deep learning models.
Entities
Institutions
- arXiv