Polyadic In-memory Computing based on Split-gate Flash Memory for Content-based Image Retrieval

Edge deployment of content-based image retrieval (CBIR) systems is crucial for real-time visual analysis in smart applications, yet it faces challenges from persistent algorithmic redundancies and von Neumann hardware bottlenecks, leading to significant energy and latency overheads. We demonstrate a polynomial neural network (PNN) accelerator based on split-gate (SPG) Flash memory to enhance the energy efficiency of CBIR systems. The SPG Flash array realizes polyadic in-memory computing (IMC) leveraging its inherent non-linear computation characteristic. The dual processing hierarchy of our CBIR system combines single-layer PNN-based similarity computation with multi-layer PNN-based feature extraction, delivering 30× speedup and 34.1 TOPs/W energy efficiency while maintaining <1% accuracy loss with 50% parameter reduction. These results present the potential of deploying efficient similarity calculations and lightweight feature extractors for in edge device.