Efficient Vision Transformer Deployment on Google Coral Through Knowledge Distillation

With the development of AI, the models are becoming more complex and resource intensive. There is a need to create smaller but at the same time satisfactory performance models that will run on resource-constrained devices. This paper explores enhancing the efficiency of the MobileNetV2 architecture using the power of the Vision Transformer through knowledge distillation. The goal was to leverage the complex insights of the Vision Transformer to improve the capabilities of MobileNetV2. As a result, the refined model outperformed its original version by about 5% (from 81.7% to 86.5%). The model was then placed on the Coral Dev Board edge device and due to the weight adjustments, there was a slight decrease in accuracy to 84.3%. Despite this, it reduced its processing latency by more than half, from 0.0076 seconds to 0.0037 seconds, proving a valuable approach for efficient and real-time edge computing applications.