An ML specialist trains a DeepAR forecasting model on CPU-based EC2 On-Demand instances and training takes many hours. Which two actions will reduce the total training time? (Choose two.)

Replacing CPU-based instances with GPU-based instances will significantly reduce training time. Deep learning models, like DeepAR, benefit greatly from the parallel processing capabilities of GPUs, which are optimized for the matrix operations common in neural networks. Training using multiple instances in parallel (distributed training) also reduces total training time by distributing the computational load across several machines, allowing the model to learn faster. Switching to Spot Instances only affects cost, not training time. Automatic model autoscaling is for deployment (inference), not training. Starting from a pretrained model and performing incremental training can reduce the amount of training needed, but it doesn't inherently speed up the training process itself; it's a different training strategy.