Researchers at Xidian University have demonstrated a photonic spiking hashing system that uses light-based neural hardware to generate stable binary codes for fast similarity retrieval. The approach offers a promising route toward high-throughput, energy-efficient search systems for applications such as retrieval-augmented generation, large-scale multimedia retrieval, and bioinformatics sequence matching.
"Hashing retrieval works by converting complex data, such as images or text, into compact 0/1 codes that can be compared rapidly," explains the study's corresponding author Shuiying Xiang. "However, implementing this process in physical photonic hardware is challenging. When a signal lies close to the decision threshold between 0 and 1, small fluctuations in optical power or readout noise may cause bit flips, reducing retrieval reliability."
To address this challenge, the Xiang and colleagues developed a hardware-software co-designed photonic spiking hashing framework. "The system first uses a single-step spiking neural network to produce continuous pre-threshold outputs," explains Xiang. "These outputs are then converted into binary hash codes by a distributed feedback laser with a saturable absorber, or DFB-SA laser, whose nonlinear threshold response acts as a physical binarization mechanism."
Notably, rather than simply encouraging correct 0/1 decisions, the training strategy pushes the network outputs farther away from the threshold boundary, creating a larger safety margin against noise-induced bit flips.
"Photonic devices are attractive for high-speed and energy-efficient computing, but their analog nature makes robust binary generation a critical issue," says Xiang. "Our method trains the network to better match the physical thresholding behavior of the DFB-SA laser, enabling more reliable photonic hash coding."
The team then validated the framework on both image and text retrieval tasks using the MNIST and 20 Newsgroups datasets. "In hardware-software collaborative experiments, 100 class-balanced queries were selected from each dataset and encoded into 12-bit hash codes," adds Xiang. "Across all tested outputs, the DFB-SA laser produced fully correct binary codes with zero observed bit flips. The end-to-end retrieval performance reached mAP = 0.9654 on MNIST and mAP = 0.9448 on 20 Newsgroups."
Taken together, these results show that photonic spiking hardware can move beyond pattern recognition and support reliable similarity retrieval. "Future integration with larger optical computing structures and parallel readout architectures may further improve throughput and latency for next-generation intelligent retrieval systems," notes Xiang.
###
Contact the author:
Shuiying Xiang
State Key Laboratory of Integrated Service Networks, Xidian University, Xi’an, China
Email: syxiang@xidian.edu.cn
The publisher KeAi was established by Elsevier and China Science Publishing & Media Ltd to unfold quality research globally. In 2013, our focus shifted to open access publishing. We now proudly publish more than 200 world-class, open access, English language journals, spanning all scientific disciplines. Many of these are titles we publish in partnership with prestigious societies and academic institutions, such as the National Natural Science Foundation of China (NSFC).
Hardware implementation of photonic spiking hash retrieval