Skip to content

Quiz: The AI HAT+ and Neural Network Fundamentals

Test your understanding of the AI HAT+ and neural network fundamentals with these review questions.

1. What is a pretrained model?

  1. A dataset that has been tagged with the correct answer by a human or trusted process
  2. A rectangle drawn tightly around a detected object to mark its location
  3. A specialized processor built to perform neural network operations
  4. A neural network that has already completed the training process and is distributed ready to use for inference
Show Answer

The correct answer is D. A pretrained model is a neural network that has already completed the training process, using someone else's dataset and computing resources, and is distributed ready to use for inference — exactly what every AI HAT+ project in this book relies on. Option A describes a labeled dataset, option B describes a bounding box, and option C describes a neural processing unit.

Concept Tested: Pretrained Model

2. What is the key difference between a training dataset and a validation dataset?

  1. The training dataset is shown to the model during training to adjust its weights and biases, while the validation dataset is withheld from training and used afterward to check performance on unseen examples
  2. The training dataset is used only for object detection, while the validation dataset is used only for image classification
  3. The validation dataset is shown to the model during training, while the training dataset is withheld and checked afterward
  4. Training and validation datasets are two names for the exact same collection of examples
Show Answer

The correct answer is A. The training dataset is the portion of labeled data actually shown to the model during training, used to adjust its weights and biases, while the validation dataset is a separate, deliberately withheld portion used afterward to check how well the model generalizes to examples it never adjusted itself to fit. Option B invents a task restriction not in the chapter, option C reverses the two roles, and option D incorrectly treats them as identical.

Concept Tested: Validation Dataset

3. How does object detection differ from image classification, according to the chapter's comparison?

  1. Image classification finds every instance of an object and reports where it is located, while object detection assigns a single label to the whole image
  2. Object detection assigns one class label to the whole image with no location information, while image classification reports multiple bounding boxes
  3. Image classification assigns a single label to the whole image without location information, while object detection finds every instance of one or more objects and reports both what and where
  4. Image classification and object detection produce identical output for the same image
Show Answer

The correct answer is C. Image classification assigns a single label to the entire image, answering "what is the main thing in this picture?" without location information, while object detection finds every instance of one or more objects and reports both what each object is and where it is located. Option A and B both reverse these roles, and option D contradicts the chapter's own comparison table showing clearly different outputs.

Concept Tested: Image Classification

4. A neuron has two inputs, x1 = 1.0 and x2 = 0.5, with weights w1 = 3 and w2 = -2, and a bias b = 0.2. Using the weighted sum formula z = w1·x1 + w2·x2 + b, what is z?

  1. 1.2
  2. 2.2
  3. 0.2
  4. 3.7
Show Answer

The correct answer is B. Substituting the values: z = (3 × 1.0) + (-2 × 0.5) + 0.2 = 3 + (-1) + 0.2 = 2.2. This follows the exact weighted-sum formula worked through with real numbers in the chapter's own example. Options A, C, and D each result from a different arithmetic slip, such as omitting a term or misapplying the bias.

Concept Tested: Neural Network

5. A student is choosing an AI HAT+ variant for a demanding project that needs to run a large object-detection model at a high frame rate. Based on the chapter's guidance, which variant should they choose?

  1. The 13 TOPS Hailo-8L version, since a lower TOPS number is always more efficient
  2. Neither variant matters, since TOPS rating has no effect on inference speed
  3. The choice should be based solely on price, ignoring TOPS rating entirely
  4. The 26 TOPS Hailo-8 version, since it offers more inference capacity for larger models or higher frame rates
Show Answer

The correct answer is D. The chapter explains that the 26 TOPS Hailo-8 variant, at a higher price, is where a more demanding project running a larger model at a higher frame rate starts to earn its cost, since TOPS measures how much inference workload a neural processing unit can handle. Option A misapplies "lower is more efficient" backwards, option B contradicts the entire purpose of the TOPS metric, and option C ignores a directly relevant technical factor.

Concept Tested: TOPS Performance Metric

6. A model produces six candidate detections with these confidence scores: 0.91, 0.87, 0.62, 0.45, 0.30, 0.12. If a project sets its confidence threshold to 0.5, how many detections are kept?

  1. 3
  2. 2
  3. 4
  4. 6
Show Answer

The correct answer is A. A confidence threshold discards any detection whose confidence score falls below the threshold; here, 0.91, 0.87, and 0.62 are all at or above 0.5 and are kept, while 0.45, 0.30, and 0.12 fall below 0.5 and are discarded, leaving 3 detections kept. Option B undercounts, option C overcounts by including a score below the threshold, and option D would require no threshold filtering at all.

Concept Tested: Confidence Score

7. A student's model achieves 99% accuracy on its training dataset but only 61% accuracy on its validation dataset. What does this pattern most likely indicate?

  1. The model is underfit and needs more training
  2. The validation dataset was mislabeled
  3. The model is overfitting, having essentially memorized the training examples instead of learning the general pattern behind them
  4. The model's confidence scores were miscalibrated
Show Answer

The correct answer is C. The chapter's explicit warning describes this exact pattern — suspiciously perfect training accuracy paired with disappointing validation accuracy — as a strong sign of overfitting, where the model has memorized specific training examples rather than learning the general pattern behind them. Option A describes the opposite problem. Option B and D each invent an unrelated explanation the chapter does not connect to this specific accuracy gap.

Concept Tested: Overfitting

8. A security camera's AI model fails to alert anyone when an intruder is clearly visible in the frame. What type of error is this?

  1. A false positive
  2. A false negative
  3. Overfitting
  4. A labeled dataset error
Show Answer

The correct answer is B. A false negative is when a model fails to detect something that actually is present — exactly the chapter's own example of a security camera missing a person who is clearly in frame. A false positive (option A) would be the reverse, reporting a detection when nothing is there. Overfitting (option C) describes a training/validation accuracy gap, not a single missed detection, and option D describes a data-quality issue unrelated to this specific error.

Concept Tested: False Negative

9. Two candidate models are being evaluated for a safety application that detects hazards on a factory floor. Model X has a false negative rate of 8% and a false positive rate of 1%. Model Y has a false negative rate of 1% and a false positive rate of 8%. Given that missing a real hazard is far more costly than a harmless false alarm, which model is the better choice for this application, and why?

  1. Model X, because its overall accuracy is easier to calculate
  2. Model X, because a higher false positive rate is always worse than a higher false negative rate
  3. Neither model can be evaluated without knowing their TOPS rating
  4. Model Y, because a lower false negative rate means it misses fewer real hazards, even though it triggers more false alarms
Show Answer

The correct answer is D. The chapter explains that a false negative in a safety application (missing a real hazard) is often much more costly than a false positive (a harmless false alarm); Model Y's much lower false negative rate (1% vs. 8%) means it misses far fewer real hazards, making it the better fit despite triggering more false alarms. Option A ignores the actual error-cost tradeoff. Option B states the opposite priority the scenario establishes, and option C introduces an irrelevant hardware factor that doesn't affect this accuracy-based judgment.

Concept Tested: False Negative

10. A team wants to design a Pi 5-based system that detects when a delivery package has been placed on a porch and sends an alert, while minimizing false alarms from passing cars or pedestrians on a nearby sidewalk. Which combination of design choices best addresses this goal, based on concepts from this chapter?

  1. Use image classification alone, since it is simpler than object detection and requires no bounding boxes
  2. Use object detection with a class label restricted to "package," combined with a confidence threshold high enough to discard weak, uncertain detections
  3. Use model training on the Pi 5 itself to build a brand-new model from scratch for this specific porch
  4. Skip confidence scores entirely and treat every detection as equally certain
Show Answer

The correct answer is B. Object detection is needed because the system must locate a specific object within a busy scene rather than just label the whole image, and filtering to the "package" class label combined with a confidence threshold directly reduces false alarms from unrelated objects like cars or pedestrians, exactly the pattern the chapter's bounding-box-and-threshold diagram demonstrates. Option A can't isolate one object within a scene. Option C contradicts the chapter's explicit point that this book only performs inference, never training, on the Pi 5, and option D discards exactly the tool built to filter out weak, unreliable detections.

Concept Tested: Object Detection