Resource-Constrained Multisource Instance-Based Transfer Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Machine learning (ML) requires extensive data, posing challenges in resource-limited settings.
• Transfer learning (TL) addresses data scarcity but faces computational/communication hurdles and negative transfer (NT).
• Existing TL research often overlooks resource consumption when tackling NT.

Purpose of the Study:

• To maximize instance-based TL accuracy in multisource, resource-constrained environments.
• To mitigate negative transfer (NT) by considering computational and communication costs.
• To introduce a novel optimization model for efficient and accurate TL.

Main Methods:

• Developed a multisource resource-constrained optimized TL (MSOPTL) model.
• MSOPTL uses a convex combination of empirical sources and target errors.
• Enhanced generalization error bounds by incorporating Kullback-Leibler (KL) divergence as a feasibility constraint.

Main Results:

• MSOPTL effectively balances TL benefits with associated costs in resource-constrained scenarios.
• Validated the model on a neural network (NN)-based classification task.
• Demonstrated improved accuracy and reduced negative transfer.

Conclusions:

• MSOPTL offers a versatile framework for various ML methods, especially in edge AI.
• The approach successfully addresses data scarcity and computational challenges in TL.
• This work advances the practical application of TL in environments with limited resources.