Autonomous Binarized Focal Loss Enhanced Model Compression Design Using Tensor Train Decomposition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models excel at object detection but often require significant computational resources.
• Model compression and pruning techniques aim to reduce the demands of deep neural networks (DNNs).
• Class imbalance remains a challenge during DNN model compression, hindering performance.

Purpose of the Study:

• To propose a novel model compression technique that addresses the class imbalance issue.
• To introduce the Autonomous Binarized Focal Loss Enhanced Model Compression (ABFLMC) model.
• To develop a hardware architecture for accelerating model inference.

Main Methods:

• Developed the Autonomous Binarized Focal Loss Enhanced Model Compression (ABFLMC) model.
• Incorporated automatic dynamic difficulty term adjustment during training.
• Designed a novel hardware architecture for inference acceleration.

Main Results:

• The ABFLMC model demonstrated improved accuracy compared to existing methods.
• Achieved faster inference speeds, enhancing real-time applicability.
• Resulted in a smaller model size, facilitating wider deployment.

Conclusions:

• The ABFLMC model effectively tackles class imbalance in DNN model compression.
• The proposed approach offers a practical solution for deploying efficient deep learning models.
• The integrated hardware architecture further optimizes inference performance.