Temporal Spiking Generative Adversarial Networks for Heading Direction Decoding

Jiangrong Shen, Kejun Wang, Wei Gao, Jian K Liu, Qi Xu, Gang Pan, Xiaodong Chen, Huajin Tang

Primary Area: applications to neuroscience & cognitive science

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.

Keywords: Generative Spiking neural networks, Heading direction

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: The spike-based neuronal responses of the ventral intraparietal area (VIP) for different heading directions appear highly spatial and temporal dynamics in the posterior parietal cortex. The data amount of biological population level VIP neuronal response is usually relatively small due to the practical data collection difficulty, which impedes the application of the complex decoding model and even causes model overfitting. To overcome the above problem, we attempt to build the unified spike-based decoding framework with a spiking neural network (SNN) for the generative and decoding model since the SNN is biologically plausible and quite suitable for neural decoding. In this paper, we propose the temporal spiking generative adversarial networks (T-SGAN) based on a spiking transformer to generate synthetic time-series data of the neuronal response of VIP neurons, followed by the recurrent SNNs with an attention mechanism to capture the spatial and temporal dynamics and decoding the heading direction. The temporal segmentation is designed in T-SGAN to reduce the length of temporal dimension and spatial self-attention is adopted to extract associated information among VIP neurons. The experiments are conducted on the collected biological datasets from monkeys to evaluate the decoding performance of the proposed framework. Experiments show that the proposed T-SGAN successfully generates realistic synthetic data and promote decoding accuracy of recurrent SNNs up to 1.75%. The above SNN-based decoding framework could further exploit the low power consumption advantages and benefit the neuronal response decoding application.

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: 2614