TransNeural: An Enhanced-Transformer-Based Performance Pre-Validation Model for Split Learning Tasks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Network Engineering

Background:

• Digital twin networks (DTNs) show promise for pre-validating network strategies.
• Split learning (SL) faces challenges in DTNs due to unknown data distributions and resource inaccuracies.
• Current DTN methods struggle to accurately estimate SL performance metrics.

Purpose of the Study:

• To propose the TransNeural algorithm for estimating SL latency and convergence within DTN pre-validation environments.
• To address limitations in current DTN approaches for SL tasks.
• To improve the accuracy of performance estimations in distributed learning networks.

Main Methods:

• The TransNeural algorithm integrates transformers to model data similarities across devices with varying distributions.
• A neural network component automatically establishes complex relationships between SL performance and system parameters.
• The method accounts for data distributions, resource availability, dataset size, and user report deviations.

Main Results:

• TransNeural demonstrated a 9.3% improvement in latency estimation accuracy.
• Convergence estimation accuracy was enhanced by 22.4% compared to traditional methods.
• The algorithm effectively models complex interdependencies influencing SL performance.

Conclusions:

• The TransNeural algorithm significantly improves the accuracy of SL latency and convergence estimations in DTNs.
• This approach offers a more robust method for pre-validating network strategies in split learning scenarios.
• Accurate DTN-based pre-validation is essential for optimizing future distributed AI systems.