Study sets
English

Fundamental Information Technology Engineer Examination (FE) | Subject A Applied AI Evaluation, Multimedia, and Edge Processing Problems 01

1 / 80.0s

Problem 1

An automated visual-inspection model treats defective products as positive. On evaluation data, it produced 72 true positives (TP), 8 false positives (FP), 18 false negatives (FN), and 102 true negatives (TN). Use precision = TP / (TP + FP) and recall = TP / (TP + FN).

Which combination of precision and recall is correct?

View explanation

Precision is 72 / (72 + 8) = 90%, the proportion of predicted defects that are actually defective. Recall is 72 / (72 + 18) = 80%, the proportion of actual defects detected. Reversing the denominators swaps the meanings of the measures, and TN is not used directly in either formula.

Problem 2

A team is developing a customer-churn model. After comparing candidate algorithms and hyperparameters, it wants one predeployment evaluation of generalization performance on data unaffected by model selection.

Which use of training, validation, and test data is most appropriate?

View explanation

Model parameters are learned from training data, while candidate models and hyperparameters are selected using validation data. Keeping the test data outside that selection process provides a fairer final estimate of performance on unseen data. Repeatedly using test results during selection leaks information into development and can make the final estimate optimistic.

Problem 3

Image classifier A achieves 99% accuracy on training data and 72% on a fixed validation set. A simpler classifier B achieves 88% and 84% on the same two sets. An independent final test set has not yet been used.

Which next step is most appropriate for selecting a model while investigating possible overfitting?

View explanation

A's large performance gap is a sign of possible overfitting. Comparing candidates and complexity using cross-validation within the development data avoids depending on a single validation split. The selected approach is then evaluated on the unused test set. Repeated tuning against that test set would fit development decisions to it and destroy its independence.

Problem 4

A fraud-detection model predicts positive when its score is at least a decision threshold. To reduce missed fraud, only the threshold is lowered while evaluation data and scores remain fixed. No observation has a score exactly equal to either the old or new threshold.

Which statement best describes the general effect of this threshold change?

View explanation

Lowering the threshold expands the set predicted positive. It may recover actual fraud that was previously missed and thereby increase recall. It can also classify more legitimate transactions as positive, increasing false positives and potentially reducing precision. The operational costs of misses and false alarms should therefore guide the threshold choice.

Problem 5

A band-limited analog signal contains frequency components up to 7.5 kHz. Under ideal conditions, it will be sampled and reconstructed using a sampling frequency greater than twice the highest signal frequency.

Which candidate is the smallest sampling frequency that satisfies the stated condition?

View explanation

Twice the highest signal frequency is 15 kHz. Because the question requires a frequency greater than twice the maximum, the smallest qualifying candidate is 16 kHz. A value of 15 kHz is equal to, not greater than, the boundary. Practical designs also allow transition bandwidth for an anti-aliasing filter.

Problem 6

Audio is recorded for 30 seconds as uncompressed PCM at a 16 kHz sampling rate, 16 bits per sample, and two stereo channels. Ignore file headers and all other overhead, and use 1 byte = 8 bits.

What is the size of the audio data in bytes?

View explanation

The size is 16,000 samples/second × 16 bits × 2 channels × 30 seconds = 15,360,000 bits. Dividing by 8 gives 1,920,000 bytes. The 960,000-byte result omits the second stereo channel, while 15,360,000 incorrectly treats a bit count as a byte count.

Problem 7

An uncompressed color image is 1,920 pixels wide by 1,080 pixels high and uses 24 bits per pixel. Ignore headers, row padding, and metadata, and use 1 byte = 8 bits.

What is the size of the pixel data in bytes?

View explanation

The image contains 1,920 × 1,080 = 2,073,600 pixels. At 24 bits, or 3 bytes, per pixel, the size is 2,073,600 × 3 = 6,220,800 bytes. The first choice assumes one byte per pixel, while the last treats 24 bits as 24 bytes.

Problem 8

Vibration sensors on factory equipment generate a large stream of data. When dangerous vibration is detected, the equipment must stop within 50 milliseconds, and shutdown decisions must continue during temporary external-network outages. The cloud stores summaries and alarm histories for maintenance analysis.

Which processing architecture best satisfies these requirements?

View explanation

The low-latency shutdown and outage-continuity requirements favor completing detection and control near the equipment on an edge device. Sending only summaries and alarm histories retains centralized analysis while reducing continuous network traffic. Edge processing still requires device security and model maintenance; it does not make cloud services universally unnecessary.