Shunbao Zhang

Curriculum Vitae

2026-10-29T00:00:00+00:00

Shunbao Zhang (张顺宝) - zhangshb2025@lzu.edu.cn

Research interests

Robustness and generalisation of deep networks; neurodynamics-inspired optimisation; loss-landscape shaping; hyperparameter learning.

Education

MSc, Software Engineering (combined MSc-PhD track), Lanzhou University, School of Information Science and Engineering, 2025.09 - present
BSc, Software Engineering, Northwest Normal University, 2020.09 - 2024.06

Publications

See the publications page.

Y. Guan, S. Zhang, L. Wang, E. Li. Enhancing Generalisation via Cascaded Inertia SGD With Learnt Hyperparameters. CAAI Transactions on Intelligence Technology, 11(4):1019-1027, 2026.
Z. Ma, X. Zhang, S. Zhang, J. Pu, X. Lin, Q. Guo. Access strategies in mmWave cell-free network: A matching and auction theory based approach. Computer Networks, 252:110659, 2024.

Awards

Postgraduate second-class scholarship, Lanzhou University
Chinese Collegiate Computing Competition: national third prize; Gansu provincial first prize
Academic research scholarship and comprehensive scholarship, Northwest Normal University

MIEO - meta integration-enhanced optimiser

2026-09-01T00:00:00+00:00

DIESGD fixes the integral order at two, so it can only filter a limited band of sharp minima - a local view of the optimisation problem. MIEO instead stacks multiple integral orders, each reshaping the loss differently and accumulating gradient history at its own timescale.

The hard part is the coefficients. Their optimal values shift during training, a feed-forward network per order does not scale, and independent integral terms let stale information accumulate. MIEO therefore differentiates through the update rule to obtain hypergradients, synthesises their direction to suppress conflicting information, and feeds the result to a controller that updates the coefficients online.

Ongoing ablations: coefficient-update rules, how many orders actually help, and loss-landscape reconstruction on ResNet.

Enhancing Generalisation via Cascaded Inertia SGD With Learnt Hyperparameters

2026-02-18T00:00:00+00:00

TL;DR - Adding a second-order integral (cascaded inertia) term to SGD filters out sharp minima during training, and the coefficients controlling it can be learned offline by a small network rather than tuned by hand.

Problem

Flat minima correlate with better generalisation and robustness, but plain SGD and momentum methods offer no direct handle on the sharpness of the basin a model settles into. Sharpness-aware methods pay for that control with extra forward/backward passes or hand-tuned schedules.

Approach

We model parameter training as a neurodynamic system and introduce a cascaded integral inertia term. The integral operator low-passes the gradient signal, which in effect removes sharp regions of the loss surface, while a lightweight feed-forward network - trained offline on noisy gradients through a neurodynamic error equation - predicts the four coefficients that govern the update.

Results

Across image-classification and remote-sensing benchmarks the method improves on standard baselines without extra per-step cost during training.

DIESGD - cascaded inertia optimiser

2026-02-18T00:00:00+00:00

Standard momentum carries first-order history; it does not care whether the model settles into a sharp or a flat basin.

DIESGD adds a second integral term: the gradient signal is integrated twice before entering the update, which acts as a low-pass filter on the loss surface and removes the sharp components that hurt generalisation. The four coefficients weighting these terms are not hand-tuned - a small feed-forward network predicts them, trained offline on noisy gradients through a neurodynamic error equation.

Published as Enhancing Generalisation via Cascaded Inertia SGD With Learnt Hyperparameters.

Access strategies in mmWave cell-free network: A matching and auction theory based approach

2024-10-01T00:00:00+00:00

Work on access strategies for millimetre-wave cell-free networks, combining matching theory and auction mechanisms to decide user association and resource allocation.