Description

Next-generation Internet-of-Everything (IoE) networks interconnecting mobile devices, heterogeneous edge servers, and cloud data centers will host diverse intelligent applications such as autonomous driving, augmented reality (AR), real-time video analytics, smart sensing, and embedded machine learning tasks. Each application relies on complex AI models with heterogeneous requirements for latency, model quality (accuracy, robustness), energy consumption, and memory footprint. Static or one-shot deployment strategies are insufficient because network conditions, computational resources, workload demand, and data characteristics evolve dynamically.

This PhD focuses on developing efficient orchestration mechanisms to provision heterogeneous intelligent applications by leveraging Dynamic AI models with adaptive architectures (e.g., early-exit networks, stem-branch structures) that enable flexible deployment, training, and inference across the device–edge–cloud continuum. The thesis addresses the challenge of jointly optimizing energy efficiency, inference and training latency, and AI model quality through intelligent management of resource allocation, compute placement, and model configuration decisions that respond to real-time system conditions.

The work will be validated on realistic testbeds integrating heterogeneous hardware accelerators (chiplet-based), wireless channel measurements, and representative AI services spanning sensor fusion, computer vision, and edge analytics. The expected outcome is a comprehensive framework for efficient orchestration that achieves energy-efficient and low-latency deployment of dynamic AI models while maintaining high model quality across resource-constrained and heterogeneous environments.

Context and Motivation

Modern intelligent applications increasingly rely on deep neural networks and complex AI models deployed across heterogeneous computational platforms: smartphones and IoT devices with limited resources, edge servers with moderate compute capabilities, and powerful cloud data centers. The deployment problem is inherently multi-dimensional: AI models must be trained and executed across these platforms, adapted to varying resource availability, and managed while satisfying stringent requirements on latency, power consumption, and inference quality.

Key challenges include:

• Resource heterogeneity: Devices have vastly different computational power, memory, storage, and energy budgets. AI models must be adapted (compressed, partitioned, or reconfigured) on a per-device basis for both training and inference.

• Dynamic conditions: Network bandwidth, device load, energy availability, and data characteristics change continuously. Static orchestration decisions are suboptimal and may violate latency or energy constraints during training or inference phases.

• Multi-objective trade-offs: Energy consumption, training/inference latency, and model quality (accuracy, robustness) are often in conflict. For example, aggressive model compression reduces energy and latency but may degrade accuracy; conversely, full-precision models ensure quality but incur higher energy and latency costs.

• Data and system variability: Input data distributions may shift over time, model performance may degrade, and new devices, applications, or network topologies may appear dynamically.

• Distributed training and inference: Coordinating model training across multiple devices with heterogeneous data and computational capabilities, and efficiently executing inference on resource-constrained platforms, requires sophisticated orchestration.

To address these challenges, this PhD leverages Dynamic AI models—AI architectures designed with built-in adaptability (early-exit mechanisms, gated modules, variable-depth branches) that enable flexible training and inference strategies. The PhD develops orchestration mechanisms that efficiently provision intelligent applications by dynamically selecting model configurations, compute placements, and resource allocations in response to system state and application requirements.

The research program “Intelligent application orchestration in next-generation networks” provides the overarching context and research methodology. This PhD specifically focuses on the dynamic ML training, inference, and distributed orchestration pillars, with emphasis on energy-efficient and latency-aware execution of both training and inference tasks for heterogeneous intelligent applications.

Research Questions and Problem Definition

The overarching problem is:

How to develop efficient orchestration mechanisms that provision heterogeneous intelligent applications using Dynamic AI models, jointly optimizing energy efficiency, training and inference latency, and AI model quality across heterogeneous device–edge–cloud systems under dynamic changes in network conditions, resource availability, and application demand?

More concretely, the PhD will address the following research questions:

RQ1: Dynamic AI model architectures and adaptive training/inference strategies

• How can AI models with dynamic architectures (early-exit networks, stem-branch designs, gated modules) be effectively utilized for both training and inference across device, edge, and cloud platforms?

• What are the optimal trade-offs between model architecture choice, computational placement, and model compression techniques?

• How can orchestration mechanisms dynamically adjust model configurations and training/inference strategies based on real-time feedback?

RQ2: Joint optimization of training/inference placement and wireless resource allocation

• How to jointly optimize placement of distributed training tasks and inference workloads while coordinating wireless spectrum allocation?

• What is the complexity of this orchestration problem, and what efficient algorithms can solve it at scale?

• How can orchestration mechanisms adapt to changes in wireless conditions, device availability, or workload demand?

RQ3: Energy-efficient training and inference in resource-constrained environments

• How can training and inference be executed energy-efficiently while maintaining strict latency and quality constraints?

• What combination of techniques best reduces energy consumption for a given quality and latency target?

• How to profile and characterize energy–latency trade-offs across architectures and hardware platforms?

RQ4: Robustness and quality assurance in dynamic training and inference orchestration

• How can orchestration mechanisms account for variability in model performance due to hardware and network differences?

• What mechanisms ensure model quality remains within acceptable bounds under dynamic adaptation?

• How to detect and mitigate performance degradation and trigger corrective actions?

Objectives of the PhD

Objective 1: Dynamic AI Model Architectures and Adaptive Training/Inference Strategies

Specific aims:

1. Characterize dynamic AI architectures in terms of flexibility and latency–accuracy–energy trade-offs.

2. Develop methods to automatically select optimal model configurations and execution strategies.

3. Design adaptive runtime policies based on feedback from training and inference execution.

4. Implement and profile models on heterogeneous hardware accelerators.

Objective 2: Joint Orchestration of Training/Inference Placement and Wireless Resource Allocation

Specific aims:

1. Formalize orchestration as a multi-dimensional optimization problem.

2. Develop efficient algorithms with theoretical guarantees.

3. Analyze complexity and provide hardness results.

4. Implement adaptive online orchestration mechanisms.

5. Evaluate scalability across large-scale heterogeneous systems.

Objective 3: Energy-Efficient and Latency-Aware Training/Inference with Quality Assurance

Specific aims:

1. Profile energy and latency across platforms and architectures.

2. Develop techniques for energy control during training and inference.

3. Design quality monitoring and corrective mechanisms.

4. Implement comprehensive testbed prototypes.

5. Conduct extensive experimental evaluations.

6. Measure convergence time, latency, energy, accuracy, throughput, and scalability.

Related Work and State of the Art

Dynamic AI Architectures and Split Computing

Dynamic AI architectures such as early-exit and stem-branch models enable resource-aware execution and have shown strong performance in reducing latency and energy while maintaining accuracy [1–3]. However, prior work largely focuses on inference and assumes static training, limited hardware heterogeneity, or static policies.

This PhD extends these works by jointly addressing training and inference orchestration, heterogeneous accelerators, adaptive runtime policies, and wireless constraints.

Distributed Training and Federated Learning

Federated and distributed learning address data locality but struggle under non-IID data, strict latency budgets, and dynamic connectivity. This PhD develops orchestration mechanisms suitable for such challenging IoE environments.

Heterogeneous Sensor Fusion and Resource-Efficient AI

Dynamic gated neural networks enable selective activation for sensor fusion, reducing computation while preserving accuracy [4–6]. This PhD generalizes these ideas to system-wide orchestration across device–edge–cloud systems.

Optimization and Algorithms for Distributed Systems

The thesis draws on task scheduling, approximation algorithms, online optimization, wireless resource allocation, and energy-aware computing.

Methodology and Approaches

Problem Formulation and Optimization

• Formalize orchestration as combinatorial optimization problems.

• Model relationships between architecture, placement, and performance.

• Apply relaxations, decomposition, and approximation techniques.

• Provide theoretical guarantees.

Algorithm Design and Implementation

• Design real-time orchestration algorithms.

• Implement centralized and distributed controllers.

• Integrate online adaptation based on execution feedback.

• Use tools such as MAESTRO for accelerator modeling.

Experimental Testbeds and Evaluation

• Build end-to-end testbeds with heterogeneous hardware and wireless emulation.

• Implement orchestration in Python/Matlab.

• Perform extensive evaluations and baseline comparisons.

• Release code and datasets.

Expected Contributions

1. A comprehensive orchestration framework for dynamic AI training and inference.

2. Formal problem formulations with theoretical analysis.

3. Efficient and scalable orchestration algorithms.

4. Adaptive runtime policies.

5. Characterization of energy–latency–quality trade-offs.

6. Open-source prototypes, datasets, and benchmarks.

7. Publications in top-tier venues.

Fit with Doctoral School, Pole, and Research Unit

The topic aligns with “Intelligent application orchestration in next-generation networks” and fits within MathSTIC / Matisse-type research poles, leveraging expertise in dynamic ML, edge computing, optimization, and network systems.

Timeline and Milestones (Indicative)

• Months 1–4: Literature review and problem formulation

• Months 4–8: Model characterization and adaptive policies (Publication 1)

• Months 9–14: Formal orchestration problem and complexity analysis

• Months 14–18: Algorithm design and prototype implementation

• Months 16–22: Testbed integration and quality assurance (Publication 2)

• Months 22–28: Comprehensive experimental evaluation (Publication 3)

• Months 28–34: Scalability and distributed variants

• Months 31–36: Thesis writing and defense