Fairness in face recognition in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding Bias Sources in Face Recognition

Which of the following is the most common source of bias in face recognition systems?

A Using a high learning rate during training

B Applying too many convolutional layers in the neural network

C Unequal representation of demographic groups in the training data

D Using grayscale images instead of color images

💡 Hint

Think about what causes the model to perform differently on various groups.

Bias often arises because some groups have fewer examples in the training data, causing the model to learn less about them.

❓ Metrics

intermediate

Evaluating Fairness Metrics

You have a face recognition model tested on two groups: Group A and Group B. The false positive rate (FPR) for Group A is 2%, and for Group B is 10%. What does this indicate about the model's fairness?

A The model is biased against Group B because it has a higher false positive rate

B The model is fair because both FPRs are below 15%

C The model is biased against Group A because it has a lower false positive rate

D The false positive rate does not relate to fairness

💡 Hint

Higher false positive rates mean more mistakes for that group.

A higher false positive rate for Group B means the model wrongly identifies more people from Group B, indicating bias against that group.

❓ Predict Output

advanced

Output of Face Embedding Normalization

What is the output of the following Python code snippet that normalizes face embeddings?

Computer Vision

import numpy as np
embeddings = np.array([[3, 4], [0, 0], [1, 1]])
norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
normalized = embeddings / norms
print(np.round(normalized, 2))

A

[[0.6 0.8]
 [0.  0. ]
 [0.71 0.71]]

B

[[0.6 0.8]
 [nan nan]
 [0.71 0.71]]

C

[[0.6 0.8]
 [0.  0. ]
 [0.5 0.5]]

D

[[0.6 0.8]
 [nan nan]
 [0.5 0.5]]

💡 Hint

Consider what happens when dividing by zero norm.

The second embedding is [0,0], so its norm is zero, causing division by zero and resulting in NaN values.

❓ Model Choice

advanced

Choosing a Model Architecture to Reduce Bias

Which model architecture choice is best to help reduce bias in face recognition across diverse populations?

A A simple logistic regression model trained on raw pixel values

B A deep CNN trained only on a large dataset from one ethnicity

C A deep CNN trained on unbalanced data without any fairness considerations

D A shallow CNN trained on a balanced dataset with demographic labels used for fairness constraints

💡 Hint

Think about dataset balance and fairness-aware training.

Using a balanced dataset and incorporating fairness constraints helps the model learn equally across groups, reducing bias.

🔧 Debug

expert

Debugging Disparate Impact in Face Recognition

A face recognition system shows 90% accuracy on Group X and 60% accuracy on Group Y. The training data is balanced. Which of the following is the most likely cause?

A The model uses a loss function that does not penalize errors differently across groups

B The model architecture is too complex for the task

C The training data contains mislabeled images for Group X only

D The model was trained with early stopping

💡 Hint

Consider how the loss function affects learning fairness.

If the loss function treats all errors equally without group awareness, the model may still perform worse on some groups despite balanced data.