2023 NIRC PhD Graduates Reveal Cutting-Edge AI and Network Intelligence Research

Ren Pengfei, PhD

Thesis title: Deep‑Vision Hand‑Gesture Pose Estimation

Abstract: Hand gestures are a natural interaction modality; estimating 3‑D hand pose from visual input faces challenges such as self‑occlusion, finger similarity, noise, and extreme viewpoints, and relies heavily on large annotated datasets. The work improves two fundamental paradigms: (1) an iterative‑correction stacked regression network that re‑parameterizes pose into pixel‑level visual cues for multi‑stage refinement, and (2) a differentiable adaptive weighting aggregation for per‑pixel pose prediction, enabling end‑to‑end optimization. The methods have garnered 106 Google Scholar citations and 140 GitHub stars.

Contributions: A stacked regression network with pose re‑parameterization; a differentiable adaptive aggregation mechanism; extensive evaluation on CVPR, TIP, AAAI, ECCV, TCYB, TCSVT with 17 papers (4 CCF‑A, 1 top‑tier Chinese Academy of Sciences journal).

He Bo, PhD

Thesis title: Multi‑Scenario Learnable Network‑Intelligent Transmission Methods

Abstract: Rapid growth of wireless services demands ultra‑low latency, high reliability, high bandwidth, and massive connectivity. Existing mathematical models struggle with dynamic wireless traffic, prompting a self‑supervised learning framework. Deep reinforcement learning (DRL) is employed without explicit environment modeling to derive optimal policies for diverse scenarios.

Key Contributions:

Cloud‑resource orchestration for differentiated services using DRL‑based traffic‑aware virtual network function provisioning.

Distributed hop‑by‑hop routing algorithm with multi‑agent DRL and attention mechanisms for edge networks.

Multi‑path transmission with DRL‑driven congestion control and self‑attention, achieving 30‑59% accuracy gains and ~8 mm reconstruction error in hand‑object interaction.

Short‑packet aggregation strategy with heuristic multi‑flow merging and DRL‑based assembly time optimization, improving deterministic delay delivery.

Tang Wei, PhD

Thesis title: Deep‑Learning Based Graph Alignment Algorithms and Applications

Abstract: Graph alignment matches nodes across graphs, crucial for tasks like social‑network user matching and knowledge‑graph fusion. Existing methods over‑focus on local structure, lack scalability, and perform poorly under weak supervision. The dissertation proposes:

A deep graph alignment model (DGAN) combining DNN and GCN modules to enhance node representation discrimination and reduce center‑node similarity inconsistencies, achieving balanced accuracy and efficiency on benchmark datasets.

A trainable cross‑graph random walk model (CEGA) that learns inter‑graph transition probabilities, outperforming prior unsupervised methods.

A multi‑order matching neighborhood consistency (MMNC) algorithm using few pseudo‑aligned seeds to align vector‑space embeddings, surpassing state‑of‑the‑art unsupervised techniques.

A position‑enhanced entity alignment model (PEEA) that incorporates positional encoding and meta‑learning for precise detection of novel, low‑sample anomalies in weakly supervised settings.

Dong Tianjian, PhD

Thesis title: Resource Scheduling Key Technologies for Knowledge‑Defined Networks

Abstract: Knowledge‑Defined Networks (KDN) add a knowledge plane atop control and data planes to interpret network state and guide decisions. The work addresses three challenges: (1) leveraging knowledge for dynamic resource allocation via GAN‑based transfer reinforcement learning; (2) joint network slicing and routing optimization using graph‑convolutional multi‑task DRL; (3) cold‑start slice orchestration with federated meta‑learning, achieving faster convergence and lower end‑to‑end delay compared with rule‑based, from‑scratch DRL, and fine‑tuned DRL baselines.

Feng Tongtong, PhD

Thesis title: Interactive Streaming Transmission Strategies and Experience Optimization

Abstract: Emerging interactive streaming applications (VR, remote surgery, autonomous driving) require ultra‑low latency, precise bitrate switching, security, and deterministic delay. The dissertation proposes four strategies:

Frame‑level transmission with RL‑based bitrate selection and rule‑based fallback, enabling sub‑second latency control.

Variable‑length media block switching with intelligent trigger mechanisms and data‑driven I‑frame prediction for accurate bitrate transitions.

RNN‑based auto‑encoder plus meta‑learning multi‑class anomaly detection for precise identification of low‑sample, imbalanced streaming attacks.

Delivery‑rate‑oriented scheduling that combines per‑frame delay sensitivity, media element priority, and inflight bandwidth prediction to guarantee deterministic delivery.

Hao Jiachang, PhD

Thesis title: Video Content Localization Based on Natural‑Language Descriptions

Abstract: The work tackles cross‑modal semantic gaps between video and text. Contributions include: (1) a text‑aware video feature encoder that re‑calibrates channelwise features using multi‑level textual semantics, achieving state‑of‑the‑art localization; (2) a video‑shuffle training framework to mitigate temporal bias and improve generalization; (3) a coarse‑boundary‑guided fine‑grained localization method that leverages subtitles and event‑response sequences for precise segment detection.

Memon Saif, PhD

Thesis title: An Enhanced QoS‑Aware Routing Protocol with Route Stability for Wireless Body Area Networks

Abstract: Wireless Body Sensor Networks (WBSNs) support delay‑sensitive health monitoring but suffer from frequent link breaks and congestion. The dissertation proposes three schemes—Enhanced Probabilistic Route Stability (EPRS), Temperature‑Link‑Delay‑aware Routing (TLD‑RP), and Cross‑Layer Improved Route Maintenance (CLIRM)—that jointly improve end‑to‑end delay, packet loss, and route stability, demonstrating significant QoS gains for medical and non‑medical applications.