How training data shapes AI behavior in AI for Everyone - Practice Exercises

Challenge - 5 Problems

🧠 Conceptual

intermediate

How does biased training data affect AI predictions?

Imagine an AI model trained mostly on images of cats with white fur. What is the most likely effect on the AI's ability to recognize cats with black fur?

A The AI will have difficulty recognizing black cats because it saw mostly white cats during training.

B The AI will ignore fur color and only focus on cat shapes, so no effect.

C The AI will recognize black cats just as well as white cats.

D The AI will refuse to classify any cat that is not white.

💡 Hint

Think about what the AI learned from the examples it saw most often.

If the training data mostly shows white cats, the AI learns features common to white cats. It may not learn features of black cats well, so it struggles to recognize them.

🔍 Analysis

intermediate

Output of model accuracy with imbalanced training data

Consider a classification model trained on 95% class A and 5% class B data. After training, the model predicts all inputs as class A. What is the approximate accuracy on the training set?

AI for Everyone

total_samples = 1000
class_A_samples = 950
class_B_samples = 50
correct_predictions = 950  # model predicts all as class A
accuracy = correct_predictions / total_samples
print(f"Accuracy: {accuracy:.2f}")

A Accuracy: 0.05

B Accuracy: 0.95

C Accuracy: 1.00

D Accuracy: 0.50

💡 Hint

How many samples are correctly predicted if the model always predicts class A?

The model correctly predicts all 950 class A samples but misses all 50 class B samples. Accuracy = 950/1000 = 0.95.

❓ Hyperparameter

advanced

Choosing training data size for better AI behavior

You want to improve an AI model's ability to generalize to new data. Which training data strategy is best?

A Use a large, diverse dataset covering many scenarios.

B Use a very small, clean dataset to avoid noise.

C Use only data from one specific case to specialize the model.

D Use random data unrelated to the task to increase size.

💡 Hint

Think about what helps the model learn to handle different situations.

A large and diverse dataset helps the model learn many patterns and generalize well to new data. Small or narrow data limits learning, and unrelated data confuses the model.

❓ Metrics

advanced

Interpreting precision and recall with skewed training data

An AI model trained on data with 90% negative and 10% positive cases achieves 90% accuracy. However, its precision is 0.5 and recall is 0.1 for the positive class. What does this tell you about the model's behavior?

A The model is good at finding positive cases but often makes false alarms.

B The model ignores the positive class and only predicts negatives.

C The model perfectly identifies positive cases but fails on negatives.

D The model misses many positive cases and is not reliable at identifying them.

💡 Hint

Recall measures how many actual positives are found. Precision measures how many predicted positives are correct.

Low recall (0.1) means the model finds only 10% of actual positives. Precision 0.5 means half of predicted positives are correct. So the model misses many positives and is unreliable for that class.

🔍 Analysis

expert

Why does this AI model fail to learn from training data?

Given this training loop snippet, why does the model's loss not decrease?

for epoch in range(5):
    for x, y in train_loader:
        optimizer.zero_grad()
        output = model(x)
        loss = loss_fn(output, y)
        loss.backward()
        optimizer.step()

A The model(x) call is missing input data, so output is empty.

B The optimizer.step() is called before loss.backward(), so gradients are not computed.

C The loss.backward function is not called properly because parentheses are missing.

D The optimizer.zero_grad() is called after loss.backward(), so gradients accumulate.

💡 Hint

Check how functions are called in Python.

loss.backward is a function and must be called with parentheses: loss.backward(). Without parentheses, gradients are not computed, so optimizer.step() does nothing.