Refer to the Exhibit Practice Questions

Add more training data.

Why wrong: Adding data increases memory requirements and would worsen the error.

Increase the learning rate.

Why wrong: Learning rate does not affect memory usage.

Switch to a CPU.

Why wrong: CPU may still run out of memory if the problem is not addressed, and training would be slower.

The intern's role is not included in the allowed roles

Why wrong: That alone wouldn't cause denial if not explicitly denied.

The policy JSON has a syntax error

Why wrong: The JSON is valid.

The endpoint path is incorrect

Why wrong: Exhibit shows correct path.

Vanishing gradients

Why wrong: Vanishing gradients would cause slow decrease in loss.

Underfitting

Why wrong: Underfitting would have high training loss as well.

Data leakage

Why wrong: Data leakage would cause abnormally high accuracy.

Model accuracy below threshold

Why wrong: Overall accuracy is 0.98, which is high and above typical thresholds.

High latency

Why wrong: The average latency of 45 ms is acceptable for real-time applications and not a critical concern.

Low throughput

Why wrong: Throughput of 1200 requests per second is high and not an issue.

Underfitting

Why wrong: Incorrect; underfitting would have low accuracy on both training and validation.

Data augmentation failure

Why wrong: Incorrect; data augmentation would affect training data diversity, not directly indicated.

Model compression

Why wrong: Incorrect; model compression reduces model size, not shown here.

Review the fairness check settings to ensure protected attributes are still relevant.

Why wrong: Fairness is separate; drift detection triggered due to feature drift, not fairness.

Immediately retrain the model on recent data to adapt to the drift.

Why wrong: Retraining without analyzing the nature of drift may lead to suboptimal model.

Reduce the classification threshold to 0.5 to increase sensitivity.

Why wrong: Threshold change does not address feature drift.

Underfitting

Why wrong: Underfitting would show high loss on both training and validation sets.

Data leakage

Why wrong: Data leakage would cause overly optimistic performance, not divergence.

Vanishing gradient

Why wrong: No evidence of gradients vanishing; loss is decreasing on training.

Remove dropout layers

Why wrong: Removing dropout would remove a regularization technique and likely increase overfitting.

Increase learning rate to 0.01

Why wrong: Increasing the learning rate may cause the model to converge too quickly or oscillate, often worsening overfitting.

Increase units in the first dense layer to 512

Why wrong: Increasing units increases model capacity, which typically worsens overfitting.

Ensure equal error rates across groups

Why wrong: Incorrect; equal error rates refer to equality of opportunity, not demographic parity.

Ensure equal accuracy across groups

Why wrong: Incorrect; demographic parity focuses on prediction rates, not accuracy.

Ensure model interpretability

Why wrong: Incorrect; interpretability is about understanding model decisions, not fairness metric.

Retrain the model

Why wrong: Incorrect; retraining affects accuracy, not necessarily latency.

Implement caching

Why wrong: Incorrect; caching helps for repeated queries, but does not reduce per-request latency for new inputs.

Reduce model input size

Why wrong: Incorrect; reducing input size may lower accuracy and it's a trade-off, not the primary fix.

Update an existing endpoint.

Why wrong: Updating an endpoint typically requires sagemaker:UpdateEndpoint, which is not allowed.

Delete an endpoint.

Why wrong: The policy does not include sagemaker:DeleteEndpoint.

Invoke an endpoint for inference.

Why wrong: Invoking an endpoint requires sagemaker:InvokeEndpoint, which is not included.

Trigger automatic retraining using the latest data

Why wrong: Auto-retraining without analysis could introduce bad patterns.

Roll back to the previous model version immediately

Why wrong: Rollback might be unnecessary if drift is harmless.

Ignore the drift since performance metrics are stable

Why wrong: Drift often precedes performance drop; ignoring is risky.

Change the activation to tanh

Why wrong: Tanh is similar to relu but may not resolve high learning rate issues.

Add dropout after each layer

Why wrong: Dropout regularizes but does not fix optimization instability caused by high learning rate.

Increase the number of units in the first dense layer

Why wrong: Adding units increases model capacity but does not directly address fluctuating loss.

Increase the parallelism of the Spark job

Why wrong: Parallelism helps but the feature engineering workload remains in batch; moving it earlier is more effective.

Replace Apache Flink with Apache Storm for stream processing

Why wrong: Changing stream processing engine does not reduce batch workload.

Change the output format from Parquet to CSV

Why wrong: CSV is not columnar and would be slower, increasing latency.

Remove rows with null income

Why wrong: Removing 12 rows may lose data, especially if missingness is not random.

Replace nulls with 0

Why wrong: Zero income is unrealistic and would introduce bias.

Replace nulls with the mean income of the entire dataset

Why wrong: Mean may be affected by regional differences and outliers.