Usas - Sustainable Continuous Learning Framework

Overview

Usas is a sustainable continuous learning framework for edge servers that enables adaptive learning while optimizing energy consumption. The framework addresses the challenge of maintaining up-to-date machine learning models on edge devices while managing power and computational constraints.

Key Features

• Continuous Learning: Adaptive model updates based on new data streams
• Smart Storage Management: Intelligent data and model storage optimization for edge servers
• Energy-Aware Learning: Learning algorithms that consider energy consumption during training
• Model Versioning: Efficient management of multiple model versions on resource-constrained devices
• Sustainable Operations: Optimizing the learning process for long-term sustainability

Technical Details

The framework consists of several components:

1. Learning Engine: Adaptive algorithms for continuous model updates on edge servers
2. Salient Store: Smart storage system for managing data and model artifacts efficiently
3. Energy Monitor: Real-time tracking of energy consumption during learning operations
4. Resource Scheduler: Intelligent allocation of computational resources for learning tasks
5. Model Manager: Efficient versioning and deployment of updated models

Results

Our evaluations show that Usas enables continuous learning on edge servers while reducing energy consumption compared to traditional approaches. The framework maintains model accuracy while adapting to new data patterns in real-time.

Publications

• Usas: A Sustainable Continuous-Learning Framework for Edge Servers
  Cyan S. Mishra, Deeksha Chaudhary, Jack Sampson, Vijaykrishnan Narayanan
  IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2024

• Salient Store: Enabling Smart Storage for Continuous Learning Edge Servers
  Cyan S. Mishra, Jack Sampson, Vijaykrishnan Narayanan
  CoRR (arXiv), 2024

Future Work

We are currently extending Uṣa’s to support:

• Distributed edge computing scenarios with multiple devices
• Integration with renewable energy sources
• More sophisticated carbon footprint modeling
• Broader range of edge applications beyond ML inference