P186 Spatiotemporal dynamics of FitzHugh-Nagumo based reservoir computing networks for classification tasks
Oleg V. Maslennikov*1, Dmitry S. Shchapin1, Vladimir I. Nekorkin1
1Department of Nonlinear Dynamics, Gaponov-Grekhov Institute of Applied Physics of the RAS, Nizhny Novgorod, Russia
*Email: olegmaov@gmail.com
Introduction
The paradigm of computation through dynamics is highly influential within thecomputationalneuroscience community, as it elucidates how interacting neural elements give rise to specific sensory, motor, and cognitive functions[1-3]. This framework's findings are also pivotal for advancements in artificial intelligence and are of particular interest from a nonlinear dynamics perspective[4]. This paradigm is primarily based on recurrent neural networks (RNNs), which, unlike feed-forward networks, do not simply map inputs to outputs but instead rely on their intrinsic dynamic state.
Methods
One influential approach for designing and training RNNs is reservoir computing (RC), which was proposed over two decades ago[5]. RC modifies only the output weights while keeping the recurrent weights fixed. RNNs are not only models for engineering applications but also fundamental tools for understanding basic cognitive functions that emerge from brain dynamics. From a dynamical systems perspective, their performance is closely related to the basic dynamic regime.An interesting approach relies on models traditional to computational neuroscience communitysuch as spiking dynamical neurons.
Results
In this study, we investigate networks composed of coupled FitzHugh-Nagumo(FHN)neurons and examine their capabilities for classification tasks. The neurons within these networks are interconnected via fixed electricalsynapses, and the output weights are trained within the reservoir computing framework. We utilize two-feature synthetic datasets for binary classification as inputs to our RNNs, where the output units read out neural activity to indicate the class. We employ several encoding schemesincluding time-to-first-spike and rate-basedto generate spiking patterns from static two-dimensional inputs and analyze how neural dynamics influence the performance of classification tasks.We show that, the nonlinear processing capabilities of FHNneuronsenable effective handling of complex signalssuch as discrimination of linearly inseparable classes.
Discussion
The integration ofFHNneurons into reservoir computing frameworks offers a powerful approach for tackling complex computational tasks. The model's inherent nonlinear dynamics, coupled with its ability to operate near criticality, enhances the performance and robustness of RC systems.Our resultshighlighted the efficiency of FHN-based reservoirs in achieving high classification accuracy while maintaining a manageable computational load. As research progresses, the application of these biologically inspired models is expected to expand across various fields, including robotics, neurophysiology, and artificial intelligence.
Acknowledgements
This work was supported by the Russian Science Foundation, grant No 23-72-10088.
References
1.Vyas, S., Golub, M. D., Sussillo, D., & Shenoy, K. V. (2020). Computation through neural population dynamics.Annual review of neuroscience,43(1), 249-275.
2.Barak, O. (2017).Current opinion in neurobiology,46, 1-6.
3.Sussillo, D. (2014).Current opinion in neurobiology,25, 156-163.
4.Ramezanian-Panahi, M., Abrevaya, G., Gagnon-Audet, J. C., Voleti, V., Rish, I., & Dumas, G. (2022).Frontiers in artificial intelligence,5, 807406.
5.Lukoševičius, M., & Jaeger, H. (2009).Computer science review,3(3), 127-149.
