DARDA: Domain-Aware Real-Time Dynamic Adaptation of Edge-Assisted Neural Networks

Adapting DNNs to data corruptions, ensuring efficiency in dynamic and unpredictable conditions while avoiding catastrophic forgetting

Background: In the dynamic world of mobile edge computing, DNNs often face the challenge of dealing with sudden data corruptions, which can severely degrade their performance. Traditional methods for unsupervised domain adaptation at the edge are computationally expensive and may not be feasible for real-time applications. Current approaches often suffer from catastrophic forgetting, where adapting to new conditions causes the model to lose previously learned knowledge. The need for efficient, real-time adaptation methods that can handle unpredictable data corruptions while maintaining computational efficiency is critical for edge computing applications.

Technical Overview: Northeastern researchers have developed Domain-Aware Real-Time Dynamic Adaptation (DARDA), a novel approach that proactively learns latent representations of a limited number of corruption types before deployment. The system employs a lightweight adaptation mechanism that can quickly adjust to new data corruptions without requiring extensive retraining. DARDA uses domain-aware techniques to maintain performance across different corruption types while avoiding catastrophic forgetting. The approach is specifically designed for edge computing environments where computational resources are limited and real-time performance is essential.

Benefits:

• Real-time adaptation to unpredictable data corruptions.
• Energy-efficient operation without compromising performance
• Avoids catastrophic forgetting during adaptation
• Lightweight design suitable for edge computing environments
• Proactive learning approach reduces adaptation time

Application:

• Autonomous vehicle navigation systems working under varied weather conditions.
• Surveillance devices operating in changing environmental conditions
• Mobile edge computing applications requiring robust performance
• IoT devices with limited computational resources

Opportunity:

• License
• Research collaboration