Bandwidth-Efficient Distributed Neural Network Architectures With Application to Neuro-Sensor Networks.

Area of Science:

• Computer Science
• Electrical Engineering
• Biomedical Engineering

Background:

• Sensor networks face communication bandwidth constraints for distributed data processing.
• Efficient inference in resource-limited environments is crucial for applications like wearable brain-computer interfaces.

Purpose of the Study:

• To propose a design methodology for distributed neural network architectures.
• To enable efficient inference in sensor networks with limited communication bandwidth.
• To reduce power consumption in sensor network applications.

Main Methods:

• Transforming a centralized neural network into a distributed architecture with parallel branches.
• Utilizing node-specific classifiers and a dynamic compression path for bandwidth optimization.
• Validating the methodology on a motor execution task using an emulated EEG sensor network.

Main Results:

• Achieved up to a factor of 20 in bandwidth reduction.
• Achieved up to a factor of 9 in power reduction.
• Maintained minimal classification accuracy loss (up to 2%) compared to a centralized baseline.

Conclusions:

• The proposed framework effectively transforms centralized architectures into distributed, bandwidth-efficient networks.
• The methodology is suitable for low-power sensor networks and wearable brain-computer interfaces.
• The approach is applicable to various sensor network-like applications.