Efficient Neural Network Fine-Tuning via Layer Contribution Analysis

Deploying deep neural networks (DNNs) on IoT devices is growing due to privacy concerns. However, there is a contradiction between constrained computational capabilities of IoT devices and the large computational requirements of training. To improve training efficiency in DNNs, transfer learning is commonly used. This approach heavily relies on fine-tuning, which is crucial for accuracy but introduces additional computational costs. We propose an efficient fine-tuning method, Contribution-Driven Tuning (CDT), aimed at meeting accuracy requirements while speeding up the fine-tuning process. We modeled the problem of finding the optimal fine-tuning strategy to maximize accuracy by analyzing the contribution of each layer to the model performance. We utilized a solver to identify the best fine-tuning approach. Compared to meta-learning, CDT reduces the time required for fine-tuning by up to \(36\%\) , while maintaining the accuracy.