Manage AI/ML Model Training PMI-CPMAI Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A data science team is planning the training schedule for a new deep learning model. They have a fixed project deadline in four weeks. The team estimates that a single training run takes 3 days on the available GPU cluster. They need to run at least 5 experiments to tune hyperparameters, but the cluster is also shared with another team that has reserved 2 days per week for their own work. What should the team do first to ensure they meet the deadline?

   • AReduce the number of experiments to 3 to fit the schedule, without further analysis.
   • BCalculate the total GPU-days required and compare against available cluster time, then adjust the number of experiments or request additional resources.
   • CStart training immediately and monitor progress, hoping the experiments finish in time.
   • DReserve the GPU cluster for the entire four-week period, blocking the other team's access.
   Reveal Solution Discussion Select an answer first
2. 22expert · hard

   A team is training a model and observes that the loss is decreasing but very slowly. They suspect the learning rate is too low. They are using a learning rate schedule that decays over time. What should they do?

   • AReduce the learning rate further to ensure convergence.
   • BKeep the current schedule and wait longer, as the loss will eventually decrease.
   • CIncrease the initial learning rate and retrain from scratch.
   • DUse a cyclical learning rate schedule that periodically increases and decreases the learning rate.
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A team is about to train a large transformer model. They have a cluster with multiple GPUs, but the model is too large to fit on a single GPU. They need to train efficiently. What should they do?

   • ATrain on CPU instead, since it has more memory.
   • BUse a single GPU with gradient accumulation to simulate a larger batch size.
   • CReduce the model size by decreasing the number of layers, even if it hurts performance.
   • DUse model parallelism to split the model across GPUs, and also use data parallelism to increase throughput.
   Reveal Solution Discussion Select an answer first
5. 24application · medium

   A team is running multiple experiments with different hyperparameters. They need to compare the results and later audit which configuration produced the best model. What is the most effective way to track these experiments?

   • ARecord the results in a shared spreadsheet manually.
   • BSave each model file with a descriptive name, but do not log metrics.
   • CKeep a text file with notes about each experiment.
   • DUse an experiment tracking tool that automatically logs parameters, metrics, and artifacts for each run.
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   What is the purpose of coordinating hyperparameter tuning activities?

   • ATo label the training data
   • BTo systematically search for the best hyperparameter values
   • CTo deploy the model to production
   • DTo reduce the size of the training dataset
   Reveal Solution Discussion Select an answer first