Research

First Author (Under Peer Review)

• The featured transformer-in-transformer architecture with knowledge distillation enables simultaneous exploration of local and global features in natural images, facilitating faster learning from the teacher model while minimizing resource requirements.

• A novel loss function was developed to harmonize teacher and student losses, addressing the unique characteristics of hybrid-labeled images and enhancing the proposed model’s effectiveness.

• Rigorous evaluations conducted across MNIST, CIFAR10, and CIFAR100 datasets substantiate the effectiveness of the proposed approach. The empirical validation reveals noteworthy improvements in execution speed and accuracy, firmly establishing a performance benchmark.

Top