An autonomous vehicle team uses transfer learning with PyTorch and the SageMaker SDK. They need to reduce inference time for low-latency self-driving use. Which process should they use to analyze and improve model inference performance?

The correct answer leverages SageMaker Debugger for model pruning, a technique to reduce model size and complexity, directly impacting inference time. Debugger allows inspection of internal model parameters (weights, gradients, biases, activations) during training. This data is crucial for identifying and ranking filters based on their importance. Pruning low-ranking filters removes redundant parts of the neural network, leading to a smaller model that executes faster. Retraining the pruned model fine-tunes the remaining parameters to maintain accuracy. Incorrect options: CloudWatch algorithm metrics do not provide the granular, internal model insights needed for filter ranking and pruning. Adjusting hyperparameters alone might improve performance but doesn't directly address model size reduction through pruning. SageMaker Model Monitor is for post-deployment monitoring of model quality and drift, not for optimizing the model architecture itself for inference speed during development.