AI-102 Plan and manage an Azure AI solution Practice Question
You deploy a Custom Vision object detection model to classify vehicles. The model works well in good lighting but fails in low-light conditions. What is the most appropriate action?
⚠ Common exam trap
Many candidates confuse model performance tuning (threshold, iterations) with data quality issues, mistakenly believing that adjusting hyperparameters can compensate for missing training scenarios.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Add images with different lighting conditions to the training set
The core issue is a data distribution mismatch: the model was trained primarily on well-lit images and lacks exposure to low-light examples. Adding images with diverse lighting conditions directly addresses this by enriching the training dataset, enabling the model to learn robust features for low-light scenarios. This aligns with the fundamental principle that Custom Vision models are only as good as the training data they receive.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✓
Add images with different lighting conditions to the training set
Why this is correct
Low-light failure is a data coverage gap, not a model or deployment fault. Custom Vision learns only from labelled examples, so adding images captured under varied lighting lets the detector learn illumination-invariant features. This directly satisfies the stem's constraint of poor performance in low-light conditions.
- ✗
Increase the probability threshold
Why it's wrong here
Raising the probability threshold only filters borderline detections; it cannot recover vehicles the model never detects in darkness, and would worsen recall. It is tempting because thresholds tune precision/recall trade-offs on validation data, which is the right lever when false positives dominate, not when illumination limits feature extraction.
- ✗
Increase the number of training iterations
Why it's wrong here
More iterations only refine the model on the existing training images; they cannot synthesise low-light features the dataset never contained. It is tempting because iteration count genuinely governs convergence when a model is underfit, but here the gap is image-domain coverage, so augmenting or adding low-light training images is required.
- ✗
Use a domain-specific model for vehicles
Why it's wrong here
A domain-specific model narrows the classification domain to vehicles but does not change how the model handles poor illumination. The failure stems from training images lacking low-light examples. Retraining with augmented low-light images addresses the actual cause; domain models suit distinguishing similar subcategories.
Go deeper
Related to this question
About these practice questions
This AI-102 question is part of Courseiva's 761-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →
JA
Written by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
This AI-102 practice question is part of Courseiva's free Microsoft certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the AI-102 exam.