2518-1092Research result. Information technologies2518-109210.18413/2518-1092-2026-11-3-0-74360ARTIFICIAL INTELLIGENCE AND DECISION MAKING<strong>PARALLELIZATION OF COMPUTATIONS&nbsp;IN LIQUID NEURAL NETWORKS WITH PRESERVATION&nbsp;OF STABILITY GUARANTEES</strong><strong>PARALLELIZATION OF COMPUTATIONS&nbsp;IN LIQUID NEURAL NETWORKS WITH PRESERVATION&nbsp;OF STABILITY GUARANTEES</strong>MongushMaksim AndreevichMongushMaksim Andreevichmakssun2018@gmail.com202611300

Liquid neural networks (LNNs) are a class of continuous machine learning models in which the evolution of a hidden state is described by a system of ordinary differential equations. A significant advantage of LNNs is their time adaptability and the presence of strict theoretical guarantees of input-to-state stability (ISS), which ensures reliable operation under noisy, irregular, and undistributed input data. These properties make LNNs promising for application in control systems, robotics, and time series processing. Despite these advantages, the practical use of LNNs is limited by high computational complexity. Numerical methods for solving ordinary differential equations have a pronounced sequential nature, which complicates efficient parallelization of computations and leads to increased training and inference times, especially when implemented on GPUs and distributed computing platforms. This paper discusses a parallelization method for LNNs: Stable Parallel Liquid Recurrence (SPLR). The approach is based on representing the problem of computing the trajectory of a hidden state over a time interval as the solution of a system of nonlinear equations equivalent to the discretized dynamics of an LNN. An iterative fixed-point method, which allows parallel execution, is used to solve it. To maintain stability at each iteration, a stabilizing projection onto the parameter and state domain in which the ISS conditions are satisfied is introduced. Theoretical results are obtained establishing the convergence of the proposed iterative algorithm and the preservation of the input-state stability property upon transition to a parallel computing scheme. It is shown that the stabilizing projection prevents the accumulation of numerical errors and limits the growth of disturbances arising during parallel processing. Unlike existing methods for parallelizing recurrent computations, which are primarily focused on speedup, the SPLR method preserves the key advantage of LNNs: their robustness to input and numerical disturbances. The practical significance of this work lies in expanding the applicability of liquid neural networks in real-time systems and on embedded platforms with limited computing resources.

Liquid neural networks (LNNs) are a class of continuous machine learning models in which the evolution of a hidden state is described by a system of ordinary differential equations. A significant advantage of LNNs is their time adaptability and the presence of strict theoretical guarantees of input-to-state stability (ISS), which ensures reliable operation under noisy, irregular, and undistributed input data. These properties make LNNs promising for application in control systems, robotics, and time series processing. Despite these advantages, the practical use of LNNs is limited by high computational complexity. Numerical methods for solving ordinary differential equations have a pronounced sequential nature, which complicates efficient parallelization of computations and leads to increased training and inference times, especially when implemented on GPUs and distributed computing platforms. This paper discusses a parallelization method for LNNs: Stable Parallel Liquid Recurrence (SPLR). The approach is based on representing the problem of computing the trajectory of a hidden state over a time interval as the solution of a system of nonlinear equations equivalent to the discretized dynamics of an LNN. An iterative fixed-point method, which allows parallel execution, is used to solve it. To maintain stability at each iteration, a stabilizing projection onto the parameter and state domain in which the ISS conditions are satisfied is introduced. Theoretical results are obtained establishing the convergence of the proposed iterative algorithm and the preservation of the input-state stability property upon transition to a parallel computing scheme. It is shown that the stabilizing projection prevents the accumulation of numerical errors and limits the growth of disturbances arising during parallel processing. Unlike existing methods for parallelizing recurrent computations, which are primarily focused on speedup, the SPLR method preserves the key advantage of LNNs: their robustness to input and numerical disturbances. The practical significance of this work lies in expanding the applicability of liquid neural networks in real-time systems and on embedded platforms with limited computing resources.

liquid neural networksparallelizationinput-to-state stabilityneural ODErecurrent neural networksdistributed computingliquid neural networksparallelizationinput-to-state stabilityneural ODErecurrent neural networksdistributed computingСписок литературыAppleyard J., Kocisk&acute;y T., Blunsom P. Optimizing Performance of Recurrent Neural Networks on GPUs [Электронный ресурс]. arXiv:1604.01946. 2016. URL: https://arxiv.org/abs/1604.01946 (дата обращения: 25.01.2026).Barabanov N.E., Prokhorov D.V. Stability Analysis of Discrete-Time Recurrent Neural Networks // IEEE Transactions on Neural Networks. 2002. Vol. 13, No. 2. P. 292-303. DOI: 10.1109/72.991416.Bidollahkhani M., Atasoy F., Abdellatef H. LTC-SE: Expanding the Potential of Liquid Time-Constant Neural Networks for Scalable AI and Embedded Systems [Электронный ресурс]. arXiv:2304.08691. 2023. DOI: 10.48550/ARXIV.2304.08691. URL: https://arxiv.org/abs/2304.08691 (дата обращения: 25.01.2026).Chahine M., Hasani R., Lechner M. et al. Robust Flight Navigation Out-of-Distribution with Liquid Neural Networks // Science Robotics. 2023. Vol. 8. eadc8892. DOI: 10.1126/scirobotics.adc8892.Chen R.T.Q., Rubanova Yu., Bettencourt J., Duvenaud D.K. Neural Ordinary Differential Equations // Advances in Neural Information Processing Systems. 2018. Vol. 31. P. 6572-6583.Cho K., van Merri&uml;enboer B., Gulcehre C. et al. Learning Phrase Representations Using RNN Encoder-Decoder for Statistical Machine Translation // Proceedings of EMNLP 2014. 2014. P. 1724-1734.Engelken R., Wolf F., Abbott L. F. Lyapunov Spectra of Chaotic Recurrent Neural Networks // Physical Review Research. 2023. Vol. 5, No. 4. 043044. DOI: 10.1103/PhysRevResearch.5.043044.Farkas A., Kert&acute;esz G., Lovas R. Parallel and Distributed Training of Deep Neural Networks: A Brief Overview // IEEE 24th International Conference on Intelligent Engineering Systems (INES). 2020. P. 165-170.Finlay C., Jacobsen J. H., Nurbekyan L., Oberman A.M. How to Train Your Neural ODE: The World of Jacobian and Kinetic Regularization // Proceedings of the 37th International Conference on Machine Learning (ICML). 2020. P. 3154 3164.Gonzalez X., Warrington A., Smith J.T.H. et al. Towards Scalable and Stable Parallelization of Nonlinear RNNs [Электронный ресурс]. arXiv:2407.19115. 2024. URL: https://arxiv.org/abs/2407.19115 (дата обращения: 25.01.2026).Guarneros-Sandoval A., Ballesteros M., Salgado I. et al. Lyapunov Stable Learning Laws for Multilayer Recurrent Neural Networks // Neurocomputing. 2022. Vol. 491. P. 644-657. DOI: 10.1016/j.neucom.2021.12.041.Hasani R., Lechner M., Amini A. et al. Liquid Time-constant Networks // Proceedings of the AAAI Conference on Artificial Intelligence. 2021. Vol. 35, No. 9. P. 7657-7666. DOI: 10.1609/aaai.v35i9.16936.Hasani R., Lechner M., Amini A. et al. Closed-form Continuous-time Neural Networks // Nature Machine Intelligence. 2022. Vol. 4, No. 11. P. 992-1003. DOI: 10.1038/s42256-022-00556-7.Hochreiter S., Schmidhuber J. Long Short-Term Memory // Neural Computation. 1997. Vol. 9, No. 8. P. 1735-1780. DOI: 10.1162/NECO.1997.9.8.1735.Kim S., Yoo J., Kim J. et al. Simulation-Free Training of Neural ODEs on Paired Data [Электронный ресурс]. arXiv:2410.22918. 2024. URL: https://arxiv.org/abs/2410.22918 (дата обращения: 25.01.2026).Krichman M., Sontag E.D., Wang Y. Input-Output-to-State Stability // SIAM Journal on Control and Optimization. 2001. Vol. 39, No. 6. P. 1874-1928.Lechner M., Hasani R., Zimmer M. et al. Neuronal Circuit Policies [Электронный ресурс]. arXiv:1803.08554. 2018. URL: https://arxiv.org/abs/1803.08554 (дата обращения: 25.01.2026).Lechner M., Hasani R., Amini A. et al. Neural Circuit Policies Enabling Auditable Autonomy // Nature Machine Intelligence. 2020. Vol. 2, No. 10. P. 642 652. DOI: 10.1038/s42256-020-00237-3.Leng J., Zhu T., Li J. et al. Liquid Neural Network in Smart Manufacturing: A New Opportunity // Advanced Engineering Informatics. 2026. Vol. 70. Art. 104222. DOI: 10.1016/j.aei.2025.104222.Li S., Zhao Y., Varma R. et al. PyTorch Distributed: Experiences on Accelerating Data Parallel Training&nbsp;// Proceedings of the VLDB Endowment. 2020. Vol. 13, No. 12. P. 3005-3018. DOI: 10.14778/3415478.3415530.Martin E., Cundy C. Parallelizing Linear Recurrent Neural Nets Over Sequence Length [Электронный ресурс]. arXiv:1709.04057. 2017. URL: https://arxiv.org/abs/1709.04057 (дата обращения: 25.01.2026).Saab Jr. S., Fu Y., Ray A., Hauser M. A Dynamically Stabilized Recurrent Neural Network // Neural Processing Letters. 2022. Vol. 54, No. 2. P. 1195-1209. DOI: 10.1007/s11063-021-10676-7.Sontag E. D. Input-to-State Stability: Basic Concepts and Results // Nonlinear and Optimal Control Theory / eds. P. Nistri, G. Stefani. Lecture Notes in Mathematics. 2008. Vol. 1932. P. 163-220. DOI: 10.1007/978-3-540-77653-6_3.Varshney L.R., Chen B.L., Paniagua E. et al. Structural Properties of the Caenorhabditis elegans Neuronal Network // PLoS Computational Biology. 2011. Vol. 7, No. 2. e1001066. DOI: 10.1371/journal.pcbi.1001066.Wang H.T. et al. Stability Analysis of Continuous-Time Switched Neural Networks with a Time-Varying Delay // IEEE Transactions on Systems, Man, and Cybernetics: Systems. 2021. Vol. 51, No. 11. P. 7022-7033.White J.G., Southgate E., Thomson J.N., Brenner S. The Structure of the Nervous System of the Nematode Caenorhabditis elegans // Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences. 1986. Vol. 314, No. 1165. P. 1-340. DOI: 10.1098/rstb.1986.0056.Yang A., Xiong J., Raginsky M., Rosenbaum E. Input-to-State Stable Neural Ordinary Differential Equations with Applications to Transient Modeling of Circuits // Learning for Dynamics and Control (L4DC). 2022. Vol. 168. P. 663-675.Zhang H., Wang Z., Liu D. A Comprehensive Review of Stability Analysis of Continuous-Time Recurrent Neural Networks // IEEE Transactions on Neural Networks and Learning Systems. 2014. Vol. 25, No. 7. P. 1229-1262. DOI: 10.1109/TNNLS.2014.2317880.Zhu M., Chang B. Convolutional Neural Networks Combined with Runge-Kutta Methods // Neural Computing and Applications. 2023. Vol. 35. P. 1629-1643. DOI: 10.1007/s00521-022-07785-2.