EFFICIENT OPERATOR CHAIN RECOGNITION VIA LOW-RANK ADAPTATION IN DEEP LEARNING FRAMEWORKS

Abstract

This thesis presents an efficient approach for recognizing operator chains in deep learn- ing frameworks using Low-Rank Adaptation (LoRA). The increasing complexity of deep learning models has created significant computational demands, making efficient adap- tation crucial for deployment in various environments. We propose a novel application of LoRA to fine-tune Vision Transformers (ViTs) for recognizing operator chains with minimal parameter updates. Our approach achieves 92.23% accuracy while fine-tuning only 0.35% of the model pa- rameters, demonstrating both computational efficiency and high recognition performance. We implement various optimization techniques, including mixed precision training, gra- dient accumulation, and early stopping to enhance both training efficiency and model performance. Our experimental results confirm that LoRA enables a significant reduction in trainable parameters while maintaining competitive performance, making it suitable for resource-constrained environments and preserving pre-trained knowledge. The frame- work’s adaptability makes it applicable across various sequence recognition tasks beyond operator chains.