Microsoft AI-102 (Designing and Implementing a Microsoft Azure AI Solution) Practice Test 33

Question 320

Inspection images can contain both corrosion and a cracked seal, and both labels must be returned. What training format is needed?

  1. Publish the intended model iteration to a prediction resource or supported deployment
  2. Use multilabel annotations allowing several positive classes on one image
  3. Review class balance and per-class precision and recall on representative samples
  4. Assign one mutually exclusive class label to each image for multiclass training

Explanation

Correct Answer

B:

Multilabel classification models independent class presence rather than enforcing a single winner. In the factory quality-control camera, confirm the fix by retesting the affected behavior after using multilabel annotations allowing several positive classes on one image.

Incorrect Answers

A:

Training output is not automatically a callable production prediction endpoint in its own workflow, but that is a different correction from the present requirement to use multilabel annotations allowing several positive classes on one image. The reported issue would remain.

C:

That change addresses another failure mode: an imbalanced dataset can yield high overall accuracy while missing the rare important defects; it does not address this case, for which the relevant intervention is to use multilabel annotations allowing several positive classes on one image.

D:

Although useful elsewhere in the factory quality-control camera, multiclass classification predicts one category among alternatives for an entire image; the present case is different and calls for the team to use multilabel annotations allowing several positive classes on one image.

Question 321

A factory classifier's overall accuracy is high, but a rare dangerous defect may be missed. Which TWO evaluation views should be added? (Choose TWO.)

  1. Account for both scale and padding offsets when mapping detection regions
  2. Review precision and recall separately for each relevant class
  3. Compute precision as 90 divided by 100, or 90%
  4. Report the number of evaluation examples for each defect class
  5. Compute recall as 80 divided by 100, or 80%

Explanation

Correct Answers

B:

Class-specific metrics expose false alarms and missed real defects that can be concealed by a dominant easy class.

D:

Support counts disclose whether rare critical defects were represented adequately to justify confidence in aggregate performance.

Incorrect Answers

A:

That change addresses another failure mode: letterboxed preprocessing changes the coordinate frame, not just the image dimensions; it does not address this case, for which the relevant intervention is to report the number of evaluation examples for each defect class and review precision and recall separately for each relevant class.

C:

Although useful elsewhere in the vision model validation pipeline, precision is TP/(TP+FP); false negatives are not part of its denominator; the present case is different and calls for the team to report the number of evaluation examples for each defect class and review precision and recall separately for each relevant class.

E:

The denominator is all actual positives, including false negatives in its own workflow, but that is a different correction from the present requirement to report the number of evaluation examples for each defect class and review precision and recall separately for each relevant class. The reported issue would remain.

Question 322

A vision model validation pipeline must measure how many real positive objects or labels the model successfully finds. Reviewers will accept the implementation when missed-defect rates are reflected in the evaluation metric. Which approach best fulfills that requirement?

  1. Review class balance and the confusion matrix
  2. Inspect prediction confidence
  3. Evaluate precision on labeled examples
  4. Evaluate recall on labeled examples

Explanation

Correct Answer

D:

Recall compares true positives with all actual positives and is sensitive to missed instances. For this vision model validation pipeline, the intended evidence is that missed-defect rates are reflected in the evaluation metric. Precision alone can be high even if the model ignores many real defects.

Incorrect Answers

A:

Per-class evaluation and confusion analysis expose false negatives and incorrect cross-class predictions; it may be used to identify systematic mix-ups between visually similar custom categories in a vision model validation pipeline, but the current requirement is instead to measure how many real positive objects or labels the model successfully finds. That separate capability does not prove that missed-defect rates are reflected in the evaluation metric; An overall accuracy figure can hide failures on rare or business-critical classes, so selecting that alternative would not meet the stated acceptance check.

B:

Confidence scores express a model's estimated certainty for particular predictions and support tuned decision thresholds; it may be used to decide which low-certainty classifications require human review in a vision model validation pipeline, but the current requirement is instead to measure how many real positive objects or labels the model successfully finds. That separate capability does not prove that missed-defect rates are reflected in the evaluation metric; A large confidence value is not proof of correctness and cannot replace labeled validation, so selecting that alternative would not meet the stated acceptance check.

C:

Precision uses true positives relative to all predicted positives, reflecting false-positive behavior; it may be used to measure how often predicted positive objects or labels are actually correct in a vision model validation pipeline, but the current requirement is instead to measure how many real positive objects or labels the model successfully finds. That separate capability does not prove that missed-defect rates are reflected in the evaluation metric; Recall instead focuses on how many actual positives were detected, so selecting that alternative would not meet the stated acceptance check.

Question 323

A release for a vision model validation pipeline is blocked until the application can decide which low-certainty classifications require human review. The tester will verify that uncertain predictions can be sent to manual review. What should be implemented?

  1. Inspect prediction confidence
  2. Evaluate precision on labeled examples
  3. Inspect bounding-box coordinates
  4. Evaluate recall on labeled examples

Explanation

Correct Answer

A:

Confidence scores express a model's estimated certainty for particular predictions and support tuned decision thresholds. For this vision model validation pipeline, the intended evidence is that uncertain predictions can be sent to manual review. A large confidence value is not proof of correctness and cannot replace labeled validation.

Incorrect Answers

B:

Precision uses true positives relative to all predicted positives, reflecting false-positive behavior; it may be used to measure how often predicted positive objects or labels are actually correct in a vision model validation pipeline, but the current requirement is instead to decide which low-certainty classifications require human review. That separate capability does not prove that uncertain predictions can be sent to manual review; Recall instead focuses on how many actual positives were detected, so selecting that alternative would not meet the stated acceptance check.

C:

Detection responses contain location geometry that can be mapped to image coordinates for highlighting or cropping; it may be used to determine the pixel region corresponding to each detected object in a vision model validation pipeline, but the current requirement is instead to decide which low-certainty classifications require human review. That separate capability does not prove that uncertain predictions can be sent to manual review; A class-level confidence score alone cannot tell where the object was found, so selecting that alternative would not meet the stated acceptance check.

D:

Recall compares true positives with all actual positives and is sensitive to missed instances; it may be used to measure how many real positive objects or labels the model successfully finds in a vision model validation pipeline, but the current requirement is instead to decide which low-certainty classifications require human review. That separate capability does not prove that uncertain predictions can be sent to manual review; Precision alone can be high even if the model ignores many real defects, so selecting that alternative would not meet the stated acceptance check.

Question 324

A threshold change reduces false alarms but lets more actual defects pass. What trade-off is being observed?

  1. Slice validation results by lighting and camera conditions
  2. Inspect the precision-recall curve as the confidence threshold changes
  3. Assess calibration and class-specific score distributions on held-out images
  4. Check the predicted box against ground-truth annotations and the task's localization tolerance

Explanation

Correct Answer

B:

Stricter decision thresholds often improve precision at the cost of recall, depending on the model's score distribution. In the vision model validation pipeline, confirm the fix by retesting the affected behavior after inspecting the precision-recall curve as the confidence threshold changes.

Incorrect Answers

A:

A single confusion matrix aggregate can conceal systematic environmental performance gaps in its own workflow, but that is a different correction from the present requirement to inspect the precision-recall curve as the confidence threshold changes. The reported issue would remain.

C:

That change addresses another failure mode: different categories can have different score characteristics, so one global threshold may perform poorly; it does not address this case, for which the relevant intervention is to inspect the precision-recall curve as the confidence threshold changes.

D:

Although useful elsewhere in the vision model validation pipeline, class confidence alone does not establish that localization meets the required intersection-over-union standard; the present case is different and calls for the team to inspect the precision-recall curve as the confidence threshold changes.

Question 325

A test reports 90 correct defect alerts and 10 false defect alerts. What is the precision?

  1. Compute precision as 90 divided by 100, or 90%
  2. Report per-class sample counts and metrics to expose minority-class failures
  3. Treat the value as model prediction confidence, not pixel-level segmentation accuracy
  4. Account for both scale and padding offsets when mapping detection regions

Explanation

Correct Answer

A:

Precision is TP/(TP+FP); false negatives are not part of its denominator. In the vision model validation pipeline, confirm the fix by retesting the affected behavior after computing precision as 90 divided by 100, or 90%.

Incorrect Answers

B:

Although useful elsewhere in the vision model validation pipeline, class distribution determines the meaning of aggregate accuracy; minority defects can be hidden; the present case is different and calls for the team to compute precision as 90 divided by 100, or 90%.

C:

Object or class confidence refers to a predicted label/region; it is not a direct measure of correct pixels in its own workflow, but that is a different correction from the present requirement to compute precision as 90 divided by 100, or 90%. The reported issue would remain.

D:

That change addresses another failure mode: letterboxed preprocessing changes the coordinate frame, not just the image dimensions; it does not address this case, for which the relevant intervention is to compute precision as 90 divided by 100, or 90%.

Question 326

After resizing a photo with letterboxing, boxes appear shifted to the right. What is missing?

  1. Report per-class sample counts and metrics to expose minority-class failures
  2. Account for both scale and padding offsets when mapping detection regions
  3. Treat the value as model prediction confidence, not pixel-level segmentation accuracy
  4. Compute precision as 90 divided by 100, or 90%

Explanation

Correct Answer

B:

Letterboxed preprocessing changes the coordinate frame, not just the image dimensions. In the vision model validation pipeline, confirm the fix by retesting the affected behavior after accounting for both scale and padding offsets when mapping detection regions.

Incorrect Answers

A:

That change addresses another failure mode: class distribution determines the meaning of aggregate accuracy; minority defects can be hidden; it does not address this case, for which the relevant intervention is to account for both scale and padding offsets when mapping detection regions.

C:

Object or class confidence refers to a predicted label/region; it is not a direct measure of correct pixels in its own workflow, but that is a different correction from the present requirement to account for both scale and padding offsets when mapping detection regions. The reported issue would remain.

D:

Although useful elsewhere in the vision model validation pipeline, precision is TP/(TP+FP); false negatives are not part of its denominator; the present case is different and calls for the team to account for both scale and padding offsets when mapping detection regions.

Question 327

In a vision model validation pipeline, engineers must measure how often predicted positive objects or labels are actually correct. The expected evidence is that false-positive-heavy experiments receive lower precision. Which implementation is the best match?

  1. Inspect bounding-box coordinates
  2. Evaluate recall on labeled examples
  3. Evaluate precision on labeled examples
  4. Review class balance and the confusion matrix

Explanation

Correct Answer

C:

Precision uses true positives relative to all predicted positives, reflecting false-positive behavior. For this vision model validation pipeline, the intended evidence is that false-positive-heavy experiments receive lower precision. Recall instead focuses on how many actual positives were detected.

Incorrect Answers

A:

Detection responses contain location geometry that can be mapped to image coordinates for highlighting or cropping; it may be used to determine the pixel region corresponding to each detected object in a vision model validation pipeline, but the current requirement is instead to measure how often predicted positive objects or labels are actually correct. That separate capability does not prove that false-positive-heavy experiments receive lower precision; A class-level confidence score alone cannot tell where the object was found, so selecting that alternative would not meet the stated acceptance check.

B:

Recall compares true positives with all actual positives and is sensitive to missed instances; it may be used to measure how many real positive objects or labels the model successfully finds in a vision model validation pipeline, but the current requirement is instead to measure how often predicted positive objects or labels are actually correct. That separate capability does not prove that false-positive-heavy experiments receive lower precision; Precision alone can be high even if the model ignores many real defects, so selecting that alternative would not meet the stated acceptance check.

D:

Per-class evaluation and confusion analysis expose false negatives and incorrect cross-class predictions; it may be used to identify systematic mix-ups between visually similar custom categories in a vision model validation pipeline, but the current requirement is instead to measure how often predicted positive objects or labels are actually correct. That separate capability does not prove that false-positive-heavy experiments receive lower precision; An overall accuracy figure can hide failures on rare or business-critical classes, so selecting that alternative would not meet the stated acceptance check.

Question 328

A test has 80 correctly detected defects and 20 missed defects. What is recall?

  1. Compute recall as 80 divided by 100, or 80%
  2. Report per-class sample counts and metrics to expose minority-class failures
  3. Account for both scale and padding offsets when mapping detection regions
  4. Compute precision as 90 divided by 100, or 90%

Explanation

Correct Answer

A:

The denominator is all actual positives, including false negatives. In the vision model validation pipeline, confirm the fix by retesting the affected behavior after computing recall as 80 divided by 100, or 80%.

Incorrect Answers

B:

Class distribution determines the meaning of aggregate accuracy; minority defects can be hidden in its own workflow, but that is a different correction from the present requirement to compute recall as 80 divided by 100, or 80%. The reported issue would remain.

C:

That change addresses another failure mode: letterboxed preprocessing changes the coordinate frame, not just the image dimensions; it does not address this case, for which the relevant intervention is to compute recall as 80 divided by 100, or 80%.

D:

Although useful elsewhere in the vision model validation pipeline, precision is TP/(TP+FP); false negatives are not part of its denominator; the present case is different and calls for the team to compute recall as 80 divided by 100, or 80%.

Question 329

Missing a defective part can create a safety incident. Which metric directly measures missed actual positives?

  1. Measure recall on the defective-item class
  2. Choose confidence thresholds using measured precision-recall trade-offs
  3. Measure precision for the positive defect predictions
  4. Convert normalized box values using the source image width and height

Explanation

Correct Answer

A:

Recall is TP/(TP+FN); it focuses on the fraction of true defective items detected. In the vision model validation pipeline, confirm the fix by retesting the affected behavior after measuring recall on the defective-item class.

Incorrect Answers

B:

Although useful elsewhere in the vision model validation pipeline, predicted confidence is a ranking signal; an arbitrary acceptance threshold may generate too many false detections; the present case is different and calls for the team to measure recall on the defective-item class.

C:

That change addresses another failure mode: precision quantifies how often predicted positives are correct; higher precision reduces unnecessary stops; it does not address this case, for which the relevant intervention is to measure recall on the defective-item class.

D:

Normalized geometry must be scaled to the image coordinate system before rendering in its own workflow, but that is a different correction from the present requirement to measure recall on the defective-item class. The reported issue would remain.

Question 330

A manager interprets a score of 0.8 as proof that eight of ten individual pixels are correct. What should be clarified?

  1. Report per-class sample counts and metrics to expose minority-class failures
  2. Account for both scale and padding offsets when mapping detection regions
  3. Compute precision as 90 divided by 100, or 90%
  4. Treat the value as model prediction confidence, not pixel-level segmentation accuracy

Explanation

Correct Answer

D:

Object or class confidence refers to a predicted label/region; it is not a direct measure of correct pixels. In the vision model validation pipeline, confirm the fix by retesting the affected behavior after treating the value as model prediction confidence, not pixel-level segmentation accuracy.

Incorrect Answers

A:

Class distribution determines the meaning of aggregate accuracy; minority defects can be hidden in its own workflow, but that is a different correction from the present requirement to treat the value as model prediction confidence, not pixel-level segmentation accuracy. The reported issue would remain.

B:

That change addresses another failure mode: letterboxed preprocessing changes the coordinate frame, not just the image dimensions; it does not address this case, for which the relevant intervention is to treat the value as model prediction confidence, not pixel-level segmentation accuracy.

C:

Although useful elsewhere in the vision model validation pipeline, precision is TP/(TP+FP); false negatives are not part of its denominator; the present case is different and calls for the team to treat the value as model prediction confidence, not pixel-level segmentation accuracy.

Designing and Implementing a Microsoft Azure AI Solution (AI-102) Practice Tests