An autonomous vehicle company trained object-detection models with transfer learning using PyTorch in SageMaker. Vehicles have limited compute and battery; the company needs to reduce model size and runtime without substantially harming accuracy. Which approach will improve the model’s computational efficiency?

The correct answer is to use Amazon SageMaker Debugger to inspect training tensors, compute filter importance, prune low-importance filters, and retrain the model. This process, known as filter pruning, directly reduces the model's complexity by removing less significant filters, thereby decreasing model size and improving runtime efficiency while minimizing accuracy loss. SageMaker Debugger is designed to access these internal model tensors during training. Using CloudWatch metrics for inspecting internal model components is incorrect because CloudWatch is primarily for monitoring infrastructure and application metrics, not for deep inspection of model tensors like weights and gradients. Using SageMaker Ground Truth to collect more data is incorrect because while more data can improve accuracy, it does not directly address model size or runtime efficiency. Using SageMaker Model Monitor to track latency and then increasing the learning rate is incorrect. Model Monitor tracks deployed model performance, but increasing the learning rate without pruning would not reduce model size or improve efficiency; it might even destabilize training.