DeepNetQoE: Self-Adaptive QoE Optimization Framework of Deep Networks.

Area of Science:

• Artificial Intelligence
• Deep Learning
• Computer Science

Background:

• Deep learning advancements and AI development heavily rely on data size and computational power.
• Researchers often sacrifice significant computing resources for improved network model precision, leading to high consumption with uncertain outcomes.
• Limited computational resources pose challenges for achieving satisfactory deep learning model performance.

Purpose of the Study:

• To propose a self-adaptive Quality of Experience (QoE) framework, DeepNetQoE, to guide deep network training.
• To establish a balance between computational resources and model performance for optimal results.
• To maximize the 'experience value' of deep learning models.

Main Methods:

• Developed a self-adaptive QoE model linking model accuracy with training computational resources.
• Established a resource allocation model to maximize the calculated experience value.
• Conducted experiments using four network models for crowd counting to analyze experience values.

Main Results:

• The DeepNetQoE framework adaptively achieves high experience values aligned with user needs.
• Experimental results demonstrate the framework's effectiveness in guiding resource allocation for deep networks.
• The proposed model successfully balances computational resource consumption and network model performance.

Conclusions:

• DeepNetQoE provides a method to guide users in determining appropriate computational resources for network models.
• The framework enables adaptive optimization of deep learning training based on user-defined experience requirements.
• Achieving a balance between resources and performance is crucial for satisfactory deep learning outcomes.