Hardware-software co-design of slimmed optical neural networks

• DOI: 10.1145/3287624.3287720
• Publication Date: 2019-01-18T21:45:18Z
• Authors: Zheng Zhao; Derong Liu; Meng Li; Zhoufeng Ying; Lu Zhang; Biying Xu; Bei Yu; Ray T. Chen; David Z. Pan
• Keywords: Neuromorphic engineering

Optical neural network (ONN) is a neuromorphic computing hardware based on optical components. Since its first on-chip experimental demonstration, it has attracted more and research interests due to the advantages of ultra-high speed inference with low power consumption. In this work, we design novel slimmed architecture for realizing considering both software implementations. Different from originally proposed ONN singular value decomposition which results in two implementation-expensive unitary matrices, show area-efficient uses sparse tree block, single block diagonal each layer. experiments, demonstrate that by leveraging training engine, are able find comparable accuracy previous architecture, brings about flexibility using implementation. The area cost terms Mach-Zehnder interferometers, core components ONN, 15%-38% less various sizes networks.