AWS Certified AI Practitioner AIF-C01 Practice Test 2
Question 10
An operations team must process a million stored images overnight rather than respond to user uploads instantly. During a design review, someone proposes Train a new model for each uploaded image. Which technical objection is specifically relevant to this proposal?
- Audio conversion does not produce requested visual predictions
- Re-training per input is not normal inference and adds avoidable cost
- Bulk offline prediction need not maintain low per-request interactive latency
- Interactive calls add unnecessary per-request pressure when offline processing is allowed
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Train a new model for each uploaded image. Re-training per input is not normal inference and adds avoidable cost
Incorrect Answers
A:This limitation concerns Replace image inference with an audio conversion service; it does not diagnose the proposed Train a new model for each uploaded image approach.
C:Bulk offline prediction need not maintain low per-request interactive latency This explains the recommended design, but it is not the requested diagnosis of Train a new model for each uploaded image.
D:This limitation concerns Choose only synchronous low-latency requests for each image; it does not diagnose the proposed Train a new model for each uploaded image approach.
Question 11
A retail platform needs to find whether an uploaded storefront image contains a logo or a bicycle. A pilot uses this approach instead: Use Amazon Transcribe to detect objects in images. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Speech synthesis provides output audio rather than image recognition
- Audio transcription cannot label objects in the supplied photograph
- Rekognition analyzes image visual content and returns labels or detected entities
- NLP sentiment does not interpret raw image pixels
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Use Amazon Transcribe to detect objects in images. Audio transcription cannot label objects in the supplied photograph
Incorrect Answers
A:This limitation concerns Use Amazon Polly to infer visual entities from narrated text; it does not diagnose the proposed Use Amazon Transcribe to detect objects in images approach.
C:Rekognition analyzes image visual content and returns labels or detected entities This explains the recommended design, but it is not the requested diagnosis of Use Amazon Transcribe to detect objects in images.
D:This limitation concerns Use Amazon Comprehend sentiment on the image file bytes; it does not diagnose the proposed Use Amazon Transcribe to detect objects in images approach.
Question 12
A streaming company has no audience-category labels but wants to discover patterns in viewing behavior. A pilot uses this approach instead: Train supervised classification using nonexistent segment labels. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Forecasting aggregates volume rather than discovering user groups
- Language translation does not group viewing patterns by behavior
- Clustering can discover similar users without predefined class labels
- A supervised classifier requires an appropriate labeled target
Explanation
Correct Answer
D:This is the actual failure mode of the proposed approach: Train supervised classification using nonexistent segment labels. A supervised classifier requires an appropriate labeled target
Incorrect Answers
A:This limitation concerns Use time-series forecasting to produce only next week's total traffic; it does not diagnose the proposed Train supervised classification using nonexistent segment labels approach.
B:This limitation concerns Use translation of descriptions as the segmentation algorithm; it does not diagnose the proposed Train supervised classification using nonexistent segment labels approach.
C:Clustering can discover similar users without predefined class labels This explains the recommended design, but it is not the requested diagnosis of Train supervised classification using nonexistent segment labels.
Question 13
A product manager confuses generative AI with ordinary binary prediction of account fraud. During a design review, someone proposes Assume a language model is required for every tabular classification. Which technical objection is specifically relevant to this proposal?
- A binary fraud label is a prediction task; generation produces new text or other content
- Neural networks can be used for classification and regression as well as generation
- Deep learning is a family of machine learning methods
- A tabular classifier may be sufficient and cheaper for labeled fixed classes
Explanation
Correct Answer
D:This is the actual failure mode of the proposed approach: Assume a language model is required for every tabular classification. A tabular classifier may be sufficient and cheaper for labeled fixed classes
Incorrect Answers
A:A binary fraud label is a prediction task; generation produces new text or other content This explains the recommended design, but it is not the requested diagnosis of Assume a language model is required for every tabular classification.
B:This limitation concerns Treat every neural network's output as necessarily a generated paragraph; it does not diagnose the proposed Assume a language model is required for every tabular classification approach.
C:This limitation concerns Describe deep learning as unrelated to machine learning; it does not diagnose the proposed Assume a language model is required for every tabular classification approach.
Question 14
A simulated warehouse robot learns which route to choose from delayed rewards for delivery time and collisions. Before approving a rollout, an architect suggests Use supervised OCR to read every shelf label as the entire control policy. Which mismatch should be documented for this precise design?
- Clustering does not optimize sequences of robot actions against rewards
- Recognizing labels is not a learned navigation policy driven by rewards
- The agent learns actions from rewards and environmental feedback across episodes
- A price estimate does not determine route actions over time
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Use supervised OCR to read every shelf label as the entire control policy. Recognizing labels is not a learned navigation policy driven by rewards
Incorrect Answers
A:This limitation concerns Use unsupervised clustering to assign static product categories; it does not diagnose the proposed Use supervised OCR to read every shelf label as the entire control policy approach.
C:The agent learns actions from rewards and environmental feedback across episodes This explains the recommended design, but it is not the requested diagnosis of Use supervised OCR to read every shelf label as the entire control policy.
D:This limitation concerns Use standard regression to output a single shipping price; it does not diagnose the proposed Use supervised OCR to read every shelf label as the entire control policy approach.
Question 15
A data analyst receives transaction tables, customer chat text and store-camera photos for one research initiative. A pilot uses this approach instead: Classify every input as tabular merely because it is stored in S3. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Target-label availability and data structure are separate properties
- Numeric table columns are not image or video pixels
- Object-storage location does not make image pixels or natural text tabular
- Each modality has different representation and preprocessing requirements
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Classify every input as tabular merely because it is stored in S3. Object-storage location does not make image pixels or natural text tabular
Incorrect Answers
A:This limitation concerns Assume labeled and unlabeled mean the same thing as structured and unstructured; it does not diagnose the proposed Classify every input as tabular merely because it is stored in S3 approach.
B:This limitation concerns Treat every numeric transaction column as a video frame; it does not diagnose the proposed Classify every input as tabular merely because it is stored in S3 approach.
D:Each modality has different representation and preprocessing requirements This explains the recommended design, but it is not the requested diagnosis of Classify every input as tabular merely because it is stored in S3.
Question 16
A support center must show an English-language chat message in Spanish without changing its meaning. Which option BEST meets the business requirement?
- Use Amazon Polly to translate by generating speech directly
- Use Amazon Comprehend sentiment to translate the sentence
- Use Amazon Textract to alter the language of an existing string
- Use Amazon Translate for language-to-language text translation
Explanation
Correct Answer
D:Translate converts text between supported natural languages The specified scenario makes this the most appropriate approach.
Incorrect Answers
A:Polly synthesizes spoken audio from text and does not provide text translation This prevents the approach from meeting the stated requirement.
B:Sentiment classification does not produce the equivalent message in Spanish This prevents the approach from meeting the stated requirement.
C:Textract extracts content from document images rather than translating phrases This prevents the approach from meeting the stated requirement.
Question 17
A binary disease-screening model misses many true positive cases although its overall accuracy looks high. A pilot uses this approach instead: Use only accuracy without inspecting class prevalence. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Accuracy can appear high when negative cases dominate
- Recall measures recovered actual positives and reveals the misses that accuracy can obscure
- Inference spending does not count clinical positives missed
- A continuous regression metric does not directly show missed positive classifications
Explanation
Correct Answer
A:This is the actual failure mode of the proposed approach: Use only accuracy without inspecting class prevalence. Accuracy can appear high when negative cases dominate
Incorrect Answers
B:Recall measures recovered actual positives and reveals the misses that accuracy can obscure This explains the recommended design, but it is not the requested diagnosis of Use only accuracy without inspecting class prevalence.
C:This limitation concerns Use token output cost as the medical false-negative measure; it does not diagnose the proposed Use only accuracy without inspecting class prevalence approach.
D:This limitation concerns Use mean absolute error as the primary binary confusion-matrix statistic; it does not diagnose the proposed Use only accuracy without inspecting class prevalence approach.
Question 18
A model training project fails because currency fields contain strings and missing values are inconsistent. Before approving a rollout, an architect suggests Assume deployment monitoring replaces initial data preparation. Which mismatch should be documented for this precise design?
- Sampling temperature does not repair tabular training data quality
- Monitoring after deployment does not clean malformed training records
- Extra compute cannot guarantee consistent features or valid numeric inputs
- Preprocessing resolves malformed values and missing-data treatment that otherwise corrupt features
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Assume deployment monitoring replaces initial data preparation. Monitoring after deployment does not clean malformed training records
Incorrect Answers
A:This limitation concerns Tune output temperature on an existing generative model to repair a broken CSV; it does not diagnose the proposed Assume deployment monitoring replaces initial data preparation approach.
C:This limitation concerns Skip training-data validation and rely only on more expensive compute; it does not diagnose the proposed Assume deployment monitoring replaces initial data preparation approach.
D:Preprocessing resolves malformed values and missing-data treatment that otherwise corrupt features This explains the recommended design, but it is not the requested diagnosis of Assume deployment monitoring replaces initial data preparation.
Question 19
An online checkout system must return a fraud score while the customer is still waiting. Before approving a rollout, an architect suggests Use a translation pipeline as the fraud scoring endpoint. Which mismatch should be documented for this precise design?
- Translating purchase descriptions is not a fraud score
- Nightly batch results arrive after the checkout decision window
- An interactive decision requires an endpoint meeting the application's latency budget
- Model training is separate from serving each inference request
Explanation
Correct Answer
A:This is the actual failure mode of the proposed approach: Use a translation pipeline as the fraud scoring endpoint. Translating purchase descriptions is not a fraud score
Incorrect Answers
B:This limitation concerns Schedule a single nightly batch job for all checkout requests; it does not diagnose the proposed Use a translation pipeline as the fraud scoring endpoint approach.
C:An interactive decision requires an endpoint meeting the application's latency budget This explains the recommended design, but it is not the requested diagnosis of Use a translation pipeline as the fraud scoring endpoint.
D:This limitation concerns Require humans to retrain the model before every purchase; it does not diagnose the proposed Use a translation pipeline as the fraud scoring endpoint approach.
Question 20
A demand model performs nearly perfectly on training records but poorly on a new seasonal holdout set. After selecting the recommended approach, which evidence MOST DIRECTLY demonstrates that the actual required outcome has been achieved?
- A reviewed design record approves this competing approach: Skip validation because the model finished deploying successfully
- Holdout error converges with training performance on representative future-period samples
- A component test invokes this alternate approach successfully: Increase the number of training epochs without checking holdout loss
- A system inventory lists this competing approach as active: Report training-set accuracy alone as proof of future accuracy
Explanation
Correct Answer
B:Holdout error converges with training performance on representative future-period samples. This checks the business result rather than just a configured feature.
Incorrect Answers
A:Deployment completion is independent of predictive generalization Design approval is not evidence that the alternative meets the actual requirement.
C:More optimization can increase overfitting when generalization is already weak A component invocation does not validate the necessary end-to-end outcome.
D:Training fit does not measure performance on unseen seasonal samples An active setting cannot independently prove the requested business outcome.
