Shunbao Zhang
Lanzhou University · School of Artificial Intelligence
I work on why deep networks generalise and stay robust. My current focus is neurodynamics-inspired optimisers that reshape the loss landscape toward flat minima.
News
- 2026.09 — Paper submitted to IEEE Transactions on Cybernetics: improving generalization through a meta integration-enhanced mechanism.
- 2026.02 — Paper accepted at CAAI Transactions on Intelligence Technology: cascaded inertia SGD with learned hyperparameters.
- 2025.09 — Started my MSc in Software Engineering at Lanzhou University.
- 2024.10 — Paper published in Computer Networks, on access strategies for mmWave cell-free networks.
Selected publications
-
Enhancing Generalisation via Cascaded Inertia SGD With Learnt HyperparametersVol. 11, No. 4, pp. 1019-1027A second-order inertia term built from neurodynamics drives SGD toward flat minima, with its coefficients learned offline by a lightweight network.
-
Access strategies in mmWave cell-free network: A matching and auction theory based approachVol. 252, Article 110659Joint user association and resource allocation in mmWave cell-free networks, solved with matching theory and auction mechanisms.
-
Improving Generalization through a Meta Integration-Enhanced Mechanism Under reviewReplacing EMA momentum with multiple integral neurodynamics of historical gradients, plus hypergradient-updated integral coefficients.
Projects
-
A bilingual academic homepage built as flat files: Zola for the build, nginx for serving, and a deploy script that swaps symlinks atomically.ZolaTera templatesnginxLet's EncryptbashDNS
Research
-
DIESGD - cascaded inertia optimiserNeurodynamics-derived second-order inertia that steers training into flat minima.
-
MIE - meta integration-enhanced optimiserMultiple integral orders over the gradient history, with coefficients updated by a hypergradient rule. Under review at IEEE Transactions on Cybernetics.