Model Prototyping Using Notebooks (e.g., Gemini Enterprise Agent Platform Workbench and Colab Enterprise) PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 1)

Questions 1–5

1. 1foundation · easy

   A data science team uses Vertex AI Workbench user-managed notebooks. The security team requires that notebook instances cannot be created outside a specific VPC and that access to the notebooks is restricted to a specific set of users. Which two Google Cloud features should be configured to enforce these requirements?

   • ACloud Identity-Aware Proxy and Cloud DNS
   • BVPC Service Controls and IAM roles
   • COrganization policies and Cloud Audit Logs
   • DCloud NAT and Cloud VPN
   Reveal Solution Discussion Select an answer first
2. 2application · medium

   A team has successfully prototyped a model in a Workbench notebook. They now want to automate the retraining of the model on a schedule and track the model versions. What is the most efficient way to integrate the notebook prototype with Vertex AI for this purpose?

   • AKeep the notebook as the source of truth and use a cron job on the Workbench instance to run the notebook on a schedule, saving the model to Cloud Storage.
   • BUse the Vertex AI Prediction API to retrain the model by sending new data to the endpoint.
   • CManually run the notebook each time new data arrives and upload the model to Vertex AI Model Registry using the console.
   • DConvert the notebook into a Vertex AI training pipeline and schedule it using Cloud Scheduler, storing the model artifacts in Vertex AI Model Registry.
   Reveal Solution Discussion Select an answer first
4. 3application · medium

   A financial services company uses Colab Enterprise for prototyping. Data scientists often work with datasets containing personally identifiable information (PII). The compliance team mandates that PII must be encrypted at rest and that data scientists must not be able to view raw PII values in their notebooks. What should you do?

   • AUse the Cloud DLP API to de-identify the data before loading it into the notebook, and store the de-identified data in a Cloud Storage bucket with default encryption.
   • BEnable customer-managed encryption keys (CMEK) on the Colab Enterprise runtime and grant data scientists access to the key to decrypt data in the notebook.
   • CCreate a VPC-SC perimeter around the Colab Enterprise API and add the data to a restricted Cloud Storage bucket, preventing access to the raw data.
   • DInstruct data scientists to manually mask PII columns using pandas before analysis, and rely on the Colab Enterprise default encryption for data at rest.
   Reveal Solution Discussion Select an answer first
5. 4expert · hard

   A researcher is prototyping a model in a Workbench notebook using JAX. They need to train a model with a custom loss function that requires computing the gradient of a complex function. The researcher is experiencing slow training and suspects the gradient computation is the bottleneck. They want to optimize the gradient computation without changing the model architecture. What is the most effective approach?

   • AUse jax.jacfwd or jax.jacrev to compute the full Jacobian matrix instead of the gradient.
   • BUse jax.experimental.checkpoint to checkpoint the loss function and reduce memory usage during gradient computation.
   • CUse jax.vmap to vectorize the gradient computation across the batch dimension.
   • DUse jax.grad to compute the gradient and apply @jax.jit to the entire training step, including the loss function and gradient computation.
   Reveal Solution Discussion Select an answer first
6. 5application · medium

   A researcher is prototyping a custom neural network in a Workbench notebook using JAX. The training loop is slow because it runs in eager mode. The researcher wants to speed up training by compiling the training step. What is the correct way to use JAX's JIT compilation?

   • AConvert the training step to use jax.experimental.jax2tf to translate the function to TensorFlow and then use TensorFlow's XLA compiler.
   • BDecorate the training step function with @jax.jit and ensure that all arguments are JAX arrays or Python scalars.
   • CUse @jax.grad on the loss function and wrap the optimizer update in a jax.lax.scan loop to enable compilation.
   • DUse the jax.numpy.vectorize function on the training step to enable automatic vectorization and compilation.
   Reveal Solution Discussion Select an answer first