Abstract
Deep learning has achieved good results in the field of image recognition due to the key role of the optimizer in a deep learning network. In this work, the optimizers of dynamical system models are established, and the influence of parameter adjustments on the dynamic performance of the system is proposed. This is a useful supplement to the theoretical control models of optimizers. First, the system control model is derived based on the iterative formula of the optimizer, the optimizer model is expressed by differential equations, and the control equation of the optimizer is established. Second, based on the system control model of the optimizer, the phase trajectory process of the optimizer model and the influence of different hyperparameters on the system performance of the learning model are analyzed. Finally, controllers with different optimizers and different hyperparameters are used to classify the MNIST and CIFAR-10 datasets to verify the effects of different optimizers on the model learning performance and compare them with related methods. Experimental results show that selecting appropriate optimizers can accelerate the convergence speed of the model and improve the accuracy of model recognition. Furthermore, the convergence speed and performance of the stochastic gradient descent (SGD) optimizer are better than those of the stochastic gradient descent-momentum (SGD-M) and Nesterov accelerated gradient (NAG) optimizers.
Similar content being viewed by others
References
Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks. Commun ACM, 2017, 60: 84–90
Sun Y L, Yu B G. Python Machine Learning Algorithm and Practice (in Chinese). Beijing: Publishing House of Electronics Industry, 2021
Robbins H, Monro S. A stochastic approximation method. Ann Math Statist, 1951, 22: 400–407
Polyak B T. Some methods of speeding up the convergence of iteration methods. USSR Comput Math Math Phys, 1964, 4: 1–17
Nesterov Y E. A method for unconstrained convex minimization problem with the rate of convergence O(1/k2). Doklady ANSSSR, 1983, 27: 543–547
An W P, Wang H Q, Sun Q Y, et al. A PID controller approach for stochastic optimization of deep networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2018. 8522–8531
Wang H Q, Luo Y, An W P, et al. PID controller-based stochastic optimization acceleration for deep neural networks. IEEE Trans Neural Netw Learn Syst, 2020, 31: 5079–5091
Wu W, Jing X Y, Du W C, et al. Learning dynamics of gradient descent optimization in deep neural networks. Sci China Inf Sci, 2021, 64: 150102
Blum J R. Multidimensional stochastic approximation methods. Ann Math Statist, 1954, 25: 737–744
Wang Z Y, Fu Y, Huang S T. Deep Learning Through Sparse and Low-Rank Modeling (in Chinese). Beijing: China Machine Press, 2021
Lei Y W, Hu T, Li G Y, et al. Stochastic gradient descent for nonconvex learning without bounded gradient assumptions. IEEE Trans Neural Netw Learn Syst, 2020, 31: 4394–4400
Engel I, Bershad N J. A transient learning comparison of Rosenblatt, backpropagation, and LMS algorithms for a single-layer perceptron for system identification. IEEE Trans Signal Process, 1994, 42: 1247–1251
Yang H H, Amari S. Complexity issues in natural gradient descent method for training multilayer perceptrons. Neural Computation, 1998, 10: 2137–2157
Bengio Y. Learning deep architectures for AI. FNT Machine Learn, 2009, 2: 1–127
Li Y Z, Liang Y Y. Learning overparameterized neural networks via stochastic gradient descent on structured data. In: Proceedings of the 32nd International Conference on Neural Information Processing Systems, Paris, 2018. 8168–8177
Luo X, Qin W, Dong A, et al. Efficient and high-quality recommendations via momentum-incorporated parallel stochastic gradient descent-based learning. IEEE CAA J Autom Sin, 2021, 8: 402–411
Ruder S. An overview of gradient descent optimization algorithms. 2017. ArXiv:1609.04747
Ding X H, Ding G G, Zhou X X, et al. Global sparse momentum SGD for pruning very deep neural networks. In: Proceedings of the 33rd Conference on Neural Information Processing Systems, Vancouver, 2019. 1–13
Wang J Y, Tantia V, Ballas N, et al. SlowMo: improving communication-efficient distributed SGD with slow momentum. In: Proceedings of International Conference on Learning Representation, 2020. 1–27
Jarek D. SGD momentum optimizer with step estimation by online parabola model. 2019. ArXiv:1907.07063
Aleksandar B, Guy L, David B. Nesterov’s accelerated gradient and momentum as approximations to regularised update descent. In: Proceedings of International Joint Conference on Neural Networks, Anchorage, 2017. 1899–1903
Luo L C, Xiong Y H, Liu Y, et al. Adaptive gradient methods with dynamic bound of learning rate. In: Proceedings of International Conference on Learning Representation, New Orleans, 2019. 1–19
Zeyuan A Z. Katyusha: the first direct acceleration of stochastic gradient methods. J Mach Learning Res, 2017, 18: 8194–8244
Luo X, Zhou M C. Effects of extended stochastic gradient descent algorithms on improving latent factor-based recommender systems. IEEE Robot Autom Lett, 2019, 4: 618–624
Luo X, Wang D X, Zhou M C, et al. Latent factor-based recommenders relying on extended stochastic gradient descent algorithms. IEEE Trans Syst Man Cyber Syst, 2021, 51: 916–926
Lei Y W, Tang K. Learning rates for stochastic gradient descent with nonconvex objectives. IEEE Trans Pattern Anal Mach Intell, 2021, 43: 4505–4511
Amid E, Anil R, Fifty C, et al. Step-size adaptation using exponentiated gradient updates. 2022. ArXiv:2202.00145
Andrychowicz M, Denil M, Colmenarejo G S, et al. Learning to learn by gradient descent by gradient descent. In: Proceedings of Advances in Neural Information Processing Systems, 2016. 1–9
Xiong Y H, Lan L C, Chen X N, et al. Learning to schedule learning rate with graph neural networks. In: Proceedings of International Conference on Learning Representations, 2022. 1–21
Shu J, Zhu Y W, Zhao Q, et al. MLR-SNet: transferable LR schedules for heterogeneous tasks. In: Proceedings of International Conference on Learning Representations, 2021. 1–25
Alamo T, Ferramosca A, González A H, et al. A gradient-based strategy for integrating real time optimizer (RTO) with model predictive control (MPC). IFAC Proc Volumes, 2012, 45: 33–38
Chen J N, Hua C C. Adaptive full-state-constrained control of nonlinear systems with deferred constraints based on nonbarrier Lyapunov function method. IEEE Trans Cybern, 2022, 52: 7634–7642
Lee T H, Trinh H M, Park J H. Stability analysis of neural networks with time-varying delay by constructing novel Lyapunov functionals. IEEE Trans Neural Netw Learn Syst, 2018, 29: 4238–4247
Faydasicok O. A new Lyapunov functional for stability analysis of neutral-type Hopfield neural networks with multiple delays. Neural Networks, 2020, 129: 288–297
Yuan F Y, Liu Y J, Liu L, et al. Adaptive neural consensus tracking control for nonlinear multiagent systems using integral barrier Lyapunov functionals. IEEE Trans Neural Netw Learn Syst, 2021. doi: https://doi.org/10.1109/TNNLS.2021.3112763
Wang Z L, Wang S K, Chen G S, et al. MATLAB/Simulink and control system simulation (in Chinese). Beijing: Publishing House of Electronics Industry, 2019
Bhaya A, Kaszkurewicz E. Control Perspectives on Numerical Algorithms and Matrix Problems. Philadelphia: Society for Industrial and Applied Mathematics, 2006
Chen Y Q, Wei Y H, Wang Y, et al. On the unified design of accelerated gradient descent. In: Proceedings of 2019 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, 2019
Liu L, Liu J, Hsieh C J, et al. Stochastically controlled compositional gradient for composition problems. IEEE Trans Neural Netw Learn Syst, 2021, doi: https://doi.org/10.1109/TNNLS.2021.3098222
Xu B G. Principle of Automatic Control (in Chinese). Beijing: Publishing House of Electronics Industry, 2013
Oppenheim A V, Willsky A S, Nawab S H. Signals and Systems (in Chinese). 2nd ed. Beijing: Publishing House of Electronics Industry, 2020
Department of Mathematics, Tongji University. Higher Mathematics (in Chinese). Beijing: Posts & Telecom Press, 2016
Gao G S, Yu W X. Principle of Automatic Control (in Chinese). Guangzhou: South China University of Technology Press, 2013
Lecun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition. Proc IEEE, 1998, 86: 2278–2324
Krizhevsky A, Hinton G. Learning Multiple Layers of Features From Tiny Images. Technical Report, University of Toronto, 2009
Acknowledgements
This work was supported by National Natural Science Foundation of China (Grant No. 61972160) and Natural Science Foundation of Guangdong Province (Grant No. 2021A1515012301).
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Yao, B., Li, G. & Wu, W. State space representation and phase analysis of gradient descent optimizers. Sci. China Inf. Sci. 66, 142102 (2023). https://doi.org/10.1007/s11432-022-3539-8
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s11432-022-3539-8