An AI system uses a pre-trained image classification model to detect defects in manufacturing. The team wants to deploy the model in an edge device with limited GPU memory. Which technique should they consider first?