Directly training on quantized model via gradient scale correction for edge device

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Quantization is crucial for deploying deep learning models on resource-constrained edge devices.
• Training quantized models directly on edge devices is challenging due to computational limitations and security concerns with data transfer.
• Existing methods fail to effectively train quantized neural networks on edge devices.

Purpose of the Study:

• To develop a novel training method for quantized models on edge devices that overcomes current limitations.
• To enhance the security and efficiency of adapting pre-trained models to private data on edge devices.
• To improve the predictive performance of quantized models trained directly on edge hardware.

Main Methods:

• Proposed a novel training method named Gradient Scale Correction and Activation Loss (GSC-AL).
• Introduced Gradient Scale Correction (GSC) to manage gradient and weight mismatches during quantized model training.
• Incorporated Activation Loss (AL) to mitigate quantization errors and activation truncation issues.

Main Results:

• The GSC-AL method was validated on seven datasets using ResNet18 and MobileNetV2.
• Significant improvements in model predictive performance were observed.
• Achieved a 41% increase in training accuracy compared to existing methods.

Conclusions:

• The GSC-AL method effectively enables direct training of quantized models on edge devices.
• This approach enhances model accuracy and addresses security concerns associated with data handling.
• The proposed method offers a promising solution for efficient and secure edge AI deployment.