Thursday 06 March 2025
The quest for efficient and effective data retrieval systems has been an ongoing challenge in the field of computer science. With the rapid growth of dynamic vision sensor (DVS) data, constructing a low-energy, efficient data retrieval system has become increasingly urgent. To tackle this issue, researchers have proposed a novel supervised hashing method named Spikinghash with a hierarchical lightweight structure.
Spikinghash is designed to take advantage of the unique characteristics of spiking neural networks (SNNs), which encode information through binary spikes. By deploying Spiking WaveMixer in shallow layers and Spiking Self-Attention in deeper layers, Spikinghash can effectively capture both local spatial features and global spatiotemporal information.
One of the key innovations of Spikinghash is its use of a dynamic soft similarity loss to leverage inter-class similarity differences. This loss function utilizes membrane potentials to construct a learnable similarity matrix as soft labels, which enables the network to fully capture the similarity differences between classes and compensate for information loss in SNNs.
The experimental results demonstrate that Spikinghash achieves state-of-the-art performance with low energy consumption and fewer parameters. For example, on the UCF101-DVS dataset, Spikinghash outperforms other methods by a significant margin, achieving an average mAP of 0.672 and an NDCG@100 of 0.734.
The success of Spikinghash can be attributed to its ability to efficiently extract features from DVS data using SNNs. The binary spikes generated by SNNs allow for fast and energy-efficient processing, making them ideal for applications where power consumption is a concern.
Furthermore, the hierarchical structure of Spikinghash enables it to effectively capture both local and global information. The shallow layers process spatial features, while the deeper layers integrate this information with temporal features to generate final hash codes.
The implications of Spikinghash are far-reaching, with potential applications in areas such as event-based computer vision, action recognition, and object detection. By leveraging the unique characteristics of SNNs, Spikinghash provides a new approach to data retrieval that is both efficient and effective.
In addition to its impressive performance, Spikinghash also offers several advantages over traditional hashing methods. For example, it does not require complex feature engineering or manual tuning of hyperparameters, making it easier to implement and train.
Cite this article: “Efficient Data Retrieval with Spikinghash: A Novel Supervised Hashing Method for Dynamic Vision Sensor Data”, The Science Archive, 2025.
Supervised Hashing, Spiking Neural Networks, Dynamic Vision Sensor, Data Retrieval, Energy-Efficient, Hierarchical Structure, Soft Similarity Loss, Membrane Potentials, Event-Based Computer Vision, Action Recognition.
