Microsoft AI-102 (Designing and Implementing a Microsoft Azure AI Solution) Practice Test 32

Question 310

A custom vision model was trained, but prediction requests target an unpublished iteration. What release step is missing?

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

Explanation

Correct Answer

D:

Training output is not automatically a callable production prediction endpoint. In the factory quality-control camera, confirm the fix by retesting the affected behavior after publishing the intended model iteration to a prediction resource or supported deployment.

Incorrect Answers

A:

Multiclass classification predicts one category among alternatives for an entire image in its own workflow, but that is a different correction from the present requirement to publish the intended model iteration to a prediction resource or supported deployment. The reported issue would remain.

B:

Although useful elsewhere in the factory quality-control camera, an imbalanced dataset can yield high overall accuracy while missing the rare important defects; the present case is different and calls for the team to publish the intended model iteration to a prediction resource or supported deployment.

C:

That change addresses another failure mode: multilabel classification models independent class presence rather than enforcing a single winner; it does not address this case, for which the relevant intervention is to publish the intended model iteration to a prediction resource or supported deployment.

Question 311

A multiclass classifier confuses scratches and dents when both appear in the same photo. What is the modeling assumption to revisit?

  1. Review bounding-box annotation quality and small-object representation
  2. Verify the model iteration and prediction endpoint referenced by the client
  3. Standardize the annotation taxonomy before training new versions
  4. Check whether the data needs multilabel or detection rather than a single-class target

Explanation

Correct Answer

D:

If multiple defect types can coexist, one-label training may be an invalid representation of the task. In the factory quality-control camera, confirm the fix by retesting the affected behavior after checking whether the data needs multilabel or detection rather than a single-class target.

Incorrect Answers

A:

Misaligned or sparse localization labels compromise box prediction despite correct class vocabulary in its own workflow, but that is a different correction from the present requirement to check whether the data needs multilabel or detection rather than a single-class target. The reported issue would remain.

B:

That change addresses another failure mode: a retrained iteration has no effect when clients continue invoking the old published version; it does not address this case, for which the relevant intervention is to check whether the data needs multilabel or detection rather than a single-class target.

C:

Although useful elsewhere in the factory quality-control camera, inconsistent labels teach contradictory distinctions and reduce model reliability; the present case is different and calls for the team to check whether the data needs multilabel or detection rather than a single-class target.

Question 312

The detector reports correct class labels but consistently misplaces boxes on small objects. What training issue should be investigated?

  1. Verify the model iteration and prediction endpoint referenced by the client
  2. Check whether the data needs multilabel or detection rather than a single-class target
  3. Standardize the annotation taxonomy before training new versions
  4. Review bounding-box annotation quality and small-object representation

Explanation

Correct Answer

D:

Misaligned or sparse localization labels compromise box prediction despite correct class vocabulary. In the factory quality-control camera, confirm the fix by retesting the affected behavior after reviewing bounding-box annotation quality and small-object representation.

Incorrect Answers

A:

Although useful elsewhere in the factory quality-control camera, a retrained iteration has no effect when clients continue invoking the old published version; the present case is different and calls for the team to review bounding-box annotation quality and small-object representation.

B:

If multiple defect types can coexist, one-label training may be an invalid representation of the task in its own workflow, but that is a different correction from the present requirement to review bounding-box annotation quality and small-object representation. The reported issue would remain.

C:

That change addresses another failure mode: inconsistent labels teach contradictory distinctions and reduce model reliability; it does not address this case, for which the relevant intervention is to review bounding-box annotation quality and small-object representation.

Question 313

A team converts multilabel outputs into one argmax class and loses co-occurring defects. What change is needed?

  1. Call the prediction service through a secured backend or identity-aware gateway
  2. Preserve each valid class prediction independently instead of taking only the top class
  3. Separate correlated units across train and evaluation splits
  4. Collect representative images from the new camera setup and reevaluate or retrain

Explanation

Correct Answer

B:

An argmax conversion defeats the purpose of multilabel output when several classes may apply. In the factory quality-control camera, confirm the fix by retesting the affected behavior after preserving each valid class prediction independently instead of taking only the top class.

Incorrect Answers

A:

Exposing service credentials in a public client undermines key secrecy and access governance in its own workflow, but that is a different correction from the present requirement to preserve each valid class prediction independently instead of taking only the top class. The reported issue would remain.

C:

Although useful elsewhere in the factory quality-control camera, near-identical images in both sets leak information and exaggerate estimated generalization; the present case is different and calls for the team to preserve each valid class prediction independently instead of taking only the top class.

D:

That change addresses another failure mode: distribution shifts in scale, lighting, and viewpoint can degrade custom detectors; it does not address this case, for which the relevant intervention is to preserve each valid class prediction independently instead of taking only the top class.

Question 314

Training labels mix “scratch” and “surface scratch” for the same defect. What data-quality task is needed?

  1. Evaluate class-specific confidence thresholds against validation data
  2. Verify the model iteration and prediction endpoint referenced by the client
  3. Check whether the data needs multilabel or detection rather than a single-class target
  4. Standardize the annotation taxonomy before training new versions

Explanation

Correct Answer

D:

Inconsistent labels teach contradictory distinctions and reduce model reliability. In the factory quality-control camera, confirm the fix by retesting the affected behavior after standardizing the annotation taxonomy before training new versions.

Incorrect Answers

A:

Although useful elsewhere in the factory quality-control camera, multiple class scores need thresholding; accepting every nonzero prediction creates false positives; the present case is different and calls for the team to standardize the annotation taxonomy before training new versions.

B:

That change addresses another failure mode: a retrained iteration has no effect when clients continue invoking the old published version; it does not address this case, for which the relevant intervention is to standardize the annotation taxonomy before training new versions.

C:

If multiple defect types can coexist, one-label training may be an invalid representation of the task in its own workflow, but that is a different correction from the present requirement to standardize the annotation taxonomy before training new versions. The reported issue would remain.

Question 315

The solution brief for a factory quality-control camera requires the system to identify several independent fault categories present in the same image. A successful test confirms that one photograph can return more than one applicable class. What should the team use?

  1. Publish and consume a trained prediction endpoint
  2. Train a multiclass image classifier
  3. Train a multilabel image classifier
  4. Prepare and label representative training images

Explanation

Correct Answer

C:

Multilabel classification allows multiple class labels to apply simultaneously rather than forcing one exclusive choice. For this factory quality-control camera, the intended evidence is that one photograph can return more than one applicable class. It does not inherently produce a bounding box for each labeled fault.

Incorrect Answers

A:

Publishing exposes a trained iteration through a supported prediction interface requiring appropriate credentials; it may be used to let a manufacturing application send new images to the trained model in a factory quality-control camera, but the current requirement is instead to identify several independent fault categories present in the same image. That separate capability does not prove that one photograph can return more than one applicable class; Training a model does not automatically connect the inference call from a production system, so selecting that alternative would not meet the stated acceptance check.

B:

Multiclass classification predicts a single label from the project's mutually exclusive classes; it may be used to assign exactly one defined quality category to each whole image in a factory quality-control camera, but the current requirement is instead to identify several independent fault categories present in the same image. That separate capability does not prove that one photograph can return more than one applicable class; It does not identify the coordinates of individual defects in the photograph, so selecting that alternative would not meet the stated acceptance check.

D:

High-quality labels and representative images establish the supervised training signal for a custom vision project; it may be used to provide the custom model with accurately annotated examples of the target classes in a factory quality-control camera, but the current requirement is instead to identify several independent fault categories present in the same image. That separate capability does not prove that one photograph can return more than one applicable class; Publishing an empty project or adding only unlabeled production images will not define the target behavior, so selecting that alternative would not meet the stated acceptance check.

Question 316

A team responsible for a factory quality-control camera needs to recognize a company-specific defect and identify where it appears; the acceptance check is that predictions include bounding boxes around localized defects. Which capability should be selected?

  1. Publish and consume a trained prediction endpoint
  2. Prepare and label representative training images
  3. Train a custom object-detection model
  4. Train a multiclass image classifier

Explanation

Correct Answer

C:

Custom object detection uses labeled regions to learn both class prediction and object location. For this factory quality-control camera, the intended evidence is that predictions include bounding boxes around localized defects. Whole-image classification lacks the required per-defect coordinates.

Incorrect Answers

A:

Publishing exposes a trained iteration through a supported prediction interface requiring appropriate credentials; it may be used to let a manufacturing application send new images to the trained model in a factory quality-control camera, but the current requirement is instead to recognize a company-specific defect and identify where it appears. That separate capability does not prove that predictions include bounding boxes around localized defects; Training a model does not automatically connect the inference call from a production system, so selecting that alternative would not meet the stated acceptance check.

B:

High-quality labels and representative images establish the supervised training signal for a custom vision project; it may be used to provide the custom model with accurately annotated examples of the target classes in a factory quality-control camera, but the current requirement is instead to recognize a company-specific defect and identify where it appears. That separate capability does not prove that predictions include bounding boxes around localized defects; Publishing an empty project or adding only unlabeled production images will not define the target behavior, so selecting that alternative would not meet the stated acceptance check.

D:

Multiclass classification predicts a single label from the project's mutually exclusive classes; it may be used to assign exactly one defined quality category to each whole image in a factory quality-control camera, but the current requirement is instead to recognize a company-specific defect and identify where it appears. That separate capability does not prove that predictions include bounding boxes around localized defects; It does not identify the coordinates of individual defects in the photograph, so selecting that alternative would not meet the stated acceptance check.

Question 317

A camera's field of view changes after training and detections degrade. What validation response fits?

  1. Separate correlated units across train and evaluation splits
  2. Preserve each valid class prediction independently instead of taking only the top class
  3. Collect labeled examples of every supported target class and retrain
  4. Collect representative images from the new camera setup and reevaluate or retrain

Explanation

Correct Answer

D:

Distribution shifts in scale, lighting, and viewpoint can degrade custom detectors. In the factory quality-control camera, confirm the fix by retesting the affected behavior after collecting representative images from the new camera setup and reevaluate or retrain.

Incorrect Answers

A:

Near-identical images in both sets leak information and exaggerate estimated generalization in its own workflow, but that is a different correction from the present requirement to collect representative images from the new camera setup and reevaluate or retrain. The reported issue would remain.

B:

That change addresses another failure mode: an argmax conversion defeats the purpose of multilabel output when several classes may apply; it does not address this case, for which the relevant intervention is to collect representative images from the new camera setup and reevaluate or retrain.

C:

Although useful elsewhere in the factory quality-control camera, a supervised closed-set classifier cannot reliably learn a class that it was never taught; the present case is different and calls for the team to collect representative images from the new camera setup and reevaluate or retrain.

Question 318

A prediction key is embedded in a browser application and exposed to every user. What should the design change?

  1. Collect representative images from the new camera setup and reevaluate or retrain
  2. Separate correlated units across train and evaluation splits
  3. Preserve each valid class prediction independently instead of taking only the top class
  4. Call the prediction service through a secured backend or identity-aware gateway

Explanation

Correct Answer

D:

Exposing service credentials in a public client undermines key secrecy and access governance. In the factory quality-control camera, confirm the fix by retesting the affected behavior after calling the prediction service through a secured backend or identity-aware gateway.

Incorrect Answers

A:

That change addresses another failure mode: distribution shifts in scale, lighting, and viewpoint can degrade custom detectors; it does not address this case, for which the relevant intervention is to call the prediction service through a secured backend or identity-aware gateway.

B:

Near-identical images in both sets leak information and exaggerate estimated generalization in its own workflow, but that is a different correction from the present requirement to call the prediction service through a secured backend or identity-aware gateway. The reported issue would remain.

C:

Although useful elsewhere in the factory quality-control camera, an argmax conversion defeats the purpose of multilabel output when several classes may apply; the present case is different and calls for the team to call the prediction service through a secured backend or identity-aware gateway.

Question 319

A release for a factory quality-control camera is blocked until the application can provide the custom model with accurately annotated examples of the target classes. The tester will verify that each training image has the expected class or region annotation. What should be implemented?

  1. Train a multiclass image classifier
  2. Publish and consume a trained prediction endpoint
  3. Train a multilabel image classifier
  4. Prepare and label representative training images
  5. Train a custom object-detection model

Explanation

Correct Answer

D:

High-quality labels and representative images establish the supervised training signal for a custom vision project. For this factory quality-control camera, the intended evidence is that each training image has the expected class or region annotation. Publishing an empty project or adding only unlabeled production images will not define the target behavior.

Incorrect Answers

A:

Multiclass classification predicts a single label from the project's mutually exclusive classes; it may be used to assign exactly one defined quality category to each whole image in a factory quality-control camera, but the current requirement is instead to provide the custom model with accurately annotated examples of the target classes. That separate capability does not prove that each training image has the expected class or region annotation; It does not identify the coordinates of individual defects in the photograph, so selecting that alternative would not meet the stated acceptance check.

B:

Publishing exposes a trained iteration through a supported prediction interface requiring appropriate credentials; it may be used to let a manufacturing application send new images to the trained model in a factory quality-control camera, but the current requirement is instead to provide the custom model with accurately annotated examples of the target classes. That separate capability does not prove that each training image has the expected class or region annotation; Training a model does not automatically connect the inference call from a production system, so selecting that alternative would not meet the stated acceptance check.

C:

Multilabel classification allows multiple class labels to apply simultaneously rather than forcing one exclusive choice; it may be used to identify several independent fault categories present in the same image in a factory quality-control camera, but the current requirement is instead to provide the custom model with accurately annotated examples of the target classes. That separate capability does not prove that each training image has the expected class or region annotation; It does not inherently produce a bounding box for each labeled fault, so selecting that alternative would not meet the stated acceptance check.

E:

Custom object detection uses labeled regions to learn both class prediction and object location; it may be used to recognize a company-specific defect and identify where it appears in a factory quality-control camera, but the current requirement is instead to provide the custom model with accurately annotated examples of the target classes. That separate capability does not prove that each training image has the expected class or region annotation; Whole-image classification lacks the required per-defect coordinates, so selecting that alternative would not meet the stated acceptance check.

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.

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