H-Direct: Homeostasis-aware Direct Spike Encoding for Deep Spiking Neural Networks

Yechan Kang, Mingyeong Seo, Jeonghee Jo, Hyun Jae Jang, Jongkil Park, Jaewook Kim, Suyoun Lee, Jinkyu Kim, Seongsik Park

Keywords: spiking neural networks; direct encoding; neuromorphic computing

Abstract: Deep spiking neural networks (SNNs) have been expected to enable energy-efficient artificial intelligence as a next-generation artificial neural network. Recently, with the development of various algorithms, such as direct spike encoding, many applications have been successfully implemented in deep SNNs. Notably, most state-of-the-art deep SNNs have greatly improved their performance by adopting direct spike encoding, which expresses input information as discrete spikes, thereby exerting substantial influence. Despite the importance of the encoding, efficient encoding methods have not been studied. As the first attempt to our knowledge, we thoroughly analyzed the conventional direct encoding. Our analysis revealed that the existing encoding restricts the training performance and efficiency due to inappropriate encoding. To address this limitation by maintaining an appropriate encoding, we introduced a concept of homeostasis to the direct spike encoding. With this concept, we presented a homeostasis-aware direct spike encoding (H-Direct), which consists of dynamic feature encoding loss, adaptive threshold, and feature diversity loss. Our experimental results demonstrate that the proposed encoding achieves higher performance and efficiency compared to conventional direct encoding across several image classification datasets on various architectures. We have validated that brain-inspired algorithms have the potential to enhance the performance and efficiency of deep SNNs.

Primary Area: other topics in machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5259