Energy-Efficient Edge and Cloud Image Classification with Multi-Reservoir Echo State Network and Data Processing...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Edge Computing

Background:

• The proliferation of Internet of Things (IoT) and cloud/edge computing necessitates efficient models for resource-constrained devices.
• Traditional deep learning models (e.g., CNNs) are computationally intensive, limiting their deployment on devices like data processing units (DPUs).
• Echo State Networks (ESNs) present a computationally efficient alternative within reservoir computing.

Purpose of the Study:

• To investigate the efficacy of ESN-based architectures, specifically Multi-Reservoir ESN (MRESN), for image classification and weather forecasting.
• To evaluate MRESN's suitability for deployment on resource-constrained hardware, such as DPUs and home stations.
• To highlight the potential of lightweight models in edge computing for efficiency and sustainability.

Main Methods:

• Utilized benchmark datasets including MNIST, FashionMnist, and CloudCast for performance evaluation.
• Implemented and assessed the Multi-Reservoir ESN (MRESN) architecture.
• Focused on evaluating computational efficiency, training times, and model adaptability.

Main Results:

• The MRESN architecture demonstrated strong performance in image classification and weather forecasting tasks.
• MRESN proved suitable for deployment on DPUs and home stations, showcasing its lightweight nature.
• The dynamic adaptability of MRESN facilitates continuous on-device training, negating the need for static pre-trained models.

Conclusions:

• Lightweight models like MRESN are crucial for efficient and sustainable cloud and edge computing applications.
• MRESN offers a versatile and high-performing alternative to traditional deep learning models in resource-constrained environments.
• This research advances efficient computing by demonstrating the practical application and scalability of MRESN architectures.