Shirong Li

Dr. Shirong Li

orcid

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
6 September 2022 in Entropy
Verified
Journal Article
Open access
PDF file
SLShirong Li
Shirong Li
XFXinlong Feng
Xinlong Feng
When PINNs solve the Navier–Stokes equations, the loss function has a gradient imbalance problem during training. It is one of the reasons why the efficiency of PINNs is limited. This paper proposes a novel method of adaptively adjusting the weights of loss terms, which can balance the gradients of each loss term during training. The weight is updated by the idea of the minmax algorithm. The neural network identifies which types of training data are harder to train and forces it to focus on those data before training the next step. Specifically, it adjusts the weight of the data that are difficult to train to maximize the objective function. On this basis, one can adjust the network parameters to minimize the objective function and do this alternately until the objective function converges. We demonstrate that the dynamic weights are monotonically non-decreasing and convergent during training. This method can not only accelerate the convergence of the loss, but also reduce the generalization error, and the computational efficiency outperformed other state-of-the-art PINNs algorithms. The validity of the method is verified by solving the forward and inverse problems of the Navier–Stokes equation.

26 February 2025 in Results in Applied Mathematics
Verified
Journal Article
Open access
SLShirong Li
Shirong Li
SLShaoyong Lai
Shaoyong Lai

11 November 2025 in Mathematics
Verified
Journal Article
Open access
PDF file
SLShirong Li
Shirong Li
HGHuan Guo
Huan Guo
MMMaliyamuguli Maimaiti
Maliyamuguli Maimaiti
SLShaoyong Lai
Shaoyong Lai
The Fornberg–Whitham equation, which contains high-order nonlinear derivatives, is widely recognized as a prominent model for describing shallow water dynamics. We explore physics-informed neural networks (PINsN) in conjunction with transfer learning technology to investigate numerical solutions for the FW equation. The proposed domain adaptation in transfer learning for PINN transforms the complex problem defined over the entire spatiotemporal domain into simpler problems defined over smaller subdomains. Training a neural network on these subdomains provides extra supervised learning data, effectively addressing optimization challenges associated with PINNs. Consequently, it enables the resolution of anisotropic and long-term predictive issues in these types of equations. Moreover, it enhances prediction accuracy and accelerates loss convergence by circumventing local optima in the specified scenarios. The method efficiently handles both forward and inverse FW equation problems, excelling in cost-effective accurate predictions for inverse problems. The efficiency and accuracy of our proposed approaches are demonstrated through examples and results.

5 November 2026 in Mathematics
Verified
Journal Article
Open access
PDF file
HGHuan Guo
Huan Guo
YLYuekai Liu
Yuekai Liu
SLShirong Li
Shirong Li
Deep convolutional neural networks have substantially advanced single-image super-resolution, or SR, but their high parameter counts and computational redundancy hinder deployment on resource-constrained edge devices. Most existing compression methods apply pruning or low-rank truncation as a post hoc step after training. This “train-then-compress” pipeline is reported in the compression literature to disrupt the spectral coherence of the learned feature manifold and to require costly fine-tuning to restore fidelity. We study an alternative, training-time approach that re-parameterizes the convolutional kernel with CANDECOMP/PARAFAC (CP) decomposition into three low-rank sublayers, optimizing directly on a rank-constrained manifold rather than compressing after convergence. We interpret the forward pass as a heuristic realization, rather than a formal equivalence, of a proximal gradient-descent–ADMM update for a tensor nuclear-norm-regularized problem, which motivates the design and its hyperparameters without invoking expensive singular value decompositions. An analysis of parameter complexity and gradient flow justifies the choice of CP over Tucker and tensor-train decompositions. Experiments on DIV2K, Kodak24, BSD100, and Set12 show that the CP module yields a 0.11 M-parameter lightweight model at parity with a same-budget reduced residual baseline (at most 0.15 dB difference on every test set), sacrificing roughly half a decibel relative to a 5.2× larger residual baseline. Applied to a large-capacity SR backbone, the same module removes 76.7% of the parameters, from 13.28 M to 3.09 M, at a PSNR cost of only 0.11–0.20 dB. These results indicate that training-time CP decomposition is a simple and parameter-efficient route to compact SR models in the extreme-compression regime.