Dynamic Image Difficulty-Aware DNN Pruning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Neural Networks (DNNs) excel in image recognition but suffer from large model sizes, hindering deployment on devices with limited resources.
• Efficient deployment of DNNs on edge devices and mobile platforms remains a significant challenge in machine learning research.

Purpose of the Study:

• To propose a novel dynamic pruning approach for Deep Neural Networks (DNNs) that adjusts pruning based on input image complexity during inference.
• To reduce the computational load and memory footprint of DNN models without compromising performance.

Main Methods:

• Developed a dynamic DNN pruning strategy that dynamically adjusts the model's complexity based on the perceived difficulty of input images.
• Evaluated the proposed method on the ImageNet dataset using various state-of-the-art DNN architectures.
• Assessed the reduction in model size and computational operations (e.g., FLOPs) achieved by the dynamic pruning technique.

Main Results:

• The dynamic pruning approach significantly reduced DNN model size and the number of operations required for inference.
• The method demonstrated effectiveness across multiple state-of-the-art DNNs without necessitating model retraining or fine-tuning.
• Achieved substantial efficiency gains, making DNNs more suitable for resource-constrained environments.

Conclusions:

• The proposed dynamic DNN pruning method offers an effective solution for creating lightweight and adaptive DNN models.
• This approach enables efficient deployment of advanced image recognition capabilities on devices with limited computational power and memory.
• Dynamic pruning presents a promising direction for future research in efficient deep learning frameworks.