AI0-001 Machine Learning and Deep Learning Practice Test 7 — 15 Questions

Question 1

Based on the exhibit, what is the likely problem with the model?

Accepted Answer

Overfitting. The exhibit shows training loss decreasing to near zero while validation loss increases after a certain point, which is a classic sign of overfitting. The model is memorizing the training data rather than learning generalizable patterns, leading to poor performance on unseen data.

Answer

Batch size too small

Answer

Learning rate too high

Answer

Underfitting

Question 2

The exhibit shows a model configuration for a classification task with 10 classes. What is wrong with this setup?

Accepted Answer

The loss function should be categorical crossentropy, not mean squared error. In a multi-class classification task with 10 classes, the correct loss function is categorical crossentropy because it measures the dissimilarity between the true probability distribution and the predicted probability distribution. Mean squared error (MSE) is designed for regression tasks and penalizes errors in a way that is not suitable for classification probabilities, leading to poor gradient behavior and slower convergence.

Answer

The metric should be precision, not accuracy

Answer

The activation should be sigmoid in hidden layers

Answer

The optimizer should be SGD, not Adam

Question 3

A data scientist is training a binary classification model to detect fraudulent transactions. The dataset is highly imbalanced with 99% legitimate and 1% fraudulent. Which evaluation metric should be prioritized to assess model performance?

Accepted Answer

F1-score. In a highly imbalanced dataset (99% legitimate, 1% fraudulent), accuracy is misleading because a model that predicts all transactions as legitimate would achieve 99% accuracy without detecting any fraud. The F1-score combines precision and recall into a single metric, making it the preferred choice for evaluating binary classification performance on imbalanced data, as it penalizes both false positives and false negatives equally.

Answer

Accuracy

Answer

Mean Squared Error

Answer

Log Loss

Question 4

A team is deploying a deep learning model for real-time image classification on edge devices with limited computational resources. Which technique would best help reduce model size and inference time without significant accuracy loss?

Accepted Answer

Model pruning and quantization. Model pruning and quantization directly reduce the number of parameters and the precision of weights (e.g., from 32-bit floats to 8-bit integers), which shrinks the model size and speeds up inference on edge devices. This technique is specifically designed to minimize computational load while preserving accuracy, making it ideal for resource-constrained environments like real-time image classification on edge hardware.

Answer

Data augmentation

Answer

Transfer learning

Answer

Ensemble learning

Question 5

A machine learning engineer notices that a linear regression model has high bias. Which action is most likely to reduce bias?

Accepted Answer

Use a more complex model, such as polynomial regression. High bias indicates that the model is too simple to capture the underlying patterns in the data, leading to underfitting. Using a more complex model, such as polynomial regression, increases the model's capacity to fit the training data better, directly addressing the underfitting issue. This is the standard approach to reduce bias in machine learning.

Answer

Reduce the number of training samples

Answer

Add L2 regularization

Answer

Apply feature scaling

Question 6

A team is developing a recommendation system for an e-commerce platform. They want to use collaborative filtering but are concerned about cold-start problems for new users. Which approach would best mitigate the cold-start problem?

Accepted Answer

Incorporate user demographic features as side information. Incorporating user demographic features as side information allows the collaborative filtering model to generate initial recommendations for new users based on their demographic profile, effectively addressing the cold-start problem. This approach uses content-based features to bootstrap the recommendation process until sufficient user interaction data is collected.

Answer

Increase the number of latent factors in matrix factorization

Answer

Use a popularity-based baseline for all recommendations

Answer

Use only item-based collaborative filtering

Question 7

A data scientist is training a deep neural network for sentiment analysis. The training loss decreases steadily but the validation loss starts to increase after 10 epochs. What is the most likely cause and best corrective action?

Accepted Answer

Overfitting; apply dropout and early stopping. The scenario describes a classic case of overfitting: the training loss decreases steadily, indicating the model is learning the training data well, but the validation loss increases after 10 epochs, meaning the model is memorizing noise and patterns specific to the training set rather than generalizing. The best corrective action is to apply dropout (which randomly drops neurons during training to reduce co-adaptation) and early stopping (which halts training when validation performance degrades), both of which are standard regularization techniques for deep neural networks.

Answer

Underfitting; increase model complexity

Answer

Vanishing gradients; use ReLU activation

Answer

Data leakage; shuffle data before splitting

Question 8

An organization wants to automate the detection of defective products on an assembly line using computer vision. They have a limited number of labeled images for defective items. Which approach would be most effective?

Accepted Answer

Use transfer learning with a pre-trained model like ResNet and fine-tune on the defect data. Transfer learning with a pre-trained model like ResNet is most effective because it leverages features learned from large datasets (e.g., ImageNet) and adapts them to the defect detection task with limited labeled data. Fine-tuning only the later layers preserves general visual features while specializing for defect classification, avoiding overfitting that would occur with a small dataset.

Answer

Use a support vector machine with handcrafted features

Answer

Train a convolutional neural network from scratch on the limited data

Answer

Synthesize additional defective images using GANs

Question 9

A machine learning engineer is troubleshooting a recurrent neural network that fails to learn long-range dependencies in sequential data. The gradients are computed using backpropagation through time. Which phenomenon is most likely occurring, and what architectural change would best address it?

Accepted Answer

Vanishing gradients; use LSTM or GRU units. In standard RNNs, backpropagation through time (BPTT) multiplies gradients across many time steps, causing them to shrink exponentially (vanishing gradients). This prevents the network from learning long-range dependencies. LSTM or GRU units introduce gating mechanisms that preserve gradient flow over many time steps, directly solving this problem.

Answer

Underfitting; increase the number of time steps

Answer

Exploding gradients; apply gradient clipping

Answer

Overfitting; reduce the number of layers

Question 10

Which TWO of the following are common activation functions used in deep neural networks?

Accepted Answer

ReLU. ReLU (Rectified Linear Unit) is a common activation function in deep neural networks because it introduces non-linearity while being computationally efficient, outputting the input directly if positive and zero otherwise. It helps mitigate the vanishing gradient problem, making it a default choice for hidden layers in many architectures.

Answer

Linear Regression

Answer

Support Vector Machine

Answer

K-means

Question 11

Which THREE of the following are techniques for handling missing data in machine learning?

Accepted Answer

Deletion of rows with missing values. Deleting rows with missing values is a straightforward technique for handling missing data, often used when the missingness is random and the dataset is large enough that removing a few rows does not significantly impact model performance. This method avoids introducing bias from imputation but can lead to loss of valuable information if too many rows are removed.

Answer

Autoencoder reconstruction

Answer

Principal Component Analysis

Question 12

Which THREE of the following are best practices for preventing overfitting in deep learning models?

Accepted Answer

L2 regularization. L2 regularization (also known as weight decay) adds a penalty term proportional to the square of the weight magnitudes to the loss function. This discourages the model from learning overly complex patterns by forcing weights to remain small, which reduces variance and helps prevent overfitting. It is a standard technique in deep learning frameworks like TensorFlow and PyTorch, where it is implemented via the `kernel_regularizer` or `weight_decay` parameter.

Answer

Increasing the number of layers

Answer

Using a larger batch size

Question 13

A hospital wants to deploy a machine learning model to predict patient readmission risk within 30 days. They have a dataset with 10,000 records, 70 features including demographics, lab results, and past admissions. The target variable is binary (readmitted or not). The data scientist trains a logistic regression model and achieves an AUC of 0.85 on the test set. However, the hospital's clinicians require interpretability of predictions to trust the model. Which action should the data scientist take to ensure the model meets the interpretability requirement while maintaining performance?

Accepted Answer

Use the logistic regression model as is, since it is inherently interpretable with coefficients. Logistic regression is a linear model whose predictions are computed as a weighted sum of input features passed through a sigmoid function. The coefficients directly represent the log-odds change per unit increase in each feature, making the model inherently interpretable without any post-hoc explanation tools. Since the model already achieves an AUC of 0.85, which meets performance requirements, no architectural change is needed. The data scientist should retain the logistic regression model and present the coefficients (and odds ratios) to clinicians to satisfy the interpretability requirement.

Answer

Reduce the number of features to 10 using PCA and retrain the logistic regression

Answer

Replace logistic regression with a random forest model and use feature importance plots

Answer

Train a deep neural network and apply LIME or SHAP for explanations

Question 14

An e-commerce company uses a gradient boosting model to forecast daily sales. Recently, the model's predictions have become less accurate, showing a significant drop in R-squared on validation data. The data scientist checks for data drift but finds no significant changes in feature distributions. The model was trained on data from the past 24 months and is retrained monthly. Upon inspecting the feature importance, the data scientist notices that the top feature 'promotion_flag' has decreased in importance over time. What is the most likely cause of the performance degradation, and what should be done?

Accepted Answer

Concept drift has occurred; retrain the model more frequently with recent data only, or use an online learning approach. The model's performance degradation, despite no data drift in feature distributions, is likely due to concept drift—the relationship between features and the target variable has changed over time. The decreasing importance of 'promotion_flag' suggests that promotions no longer influence sales as they once did. Retraining more frequently with recent data or using online learning can help the model adapt to the new concept.

Answer

The model is overfitting to historical promotions; apply more regularization

Answer

The model's hyperparameters need tuning; perform a grid search

Answer

The promotion_flag feature is leaking future information; remove it

Question 15

A financial institution uses a deep learning model for fraud detection. The model is a feedforward neural network with three hidden layers. It was trained on a balanced dataset of 100,000 transactions. During deployment, the model achieves high accuracy on the test set but the fraud detection rate (true positive rate) is only 40% while the false positive rate is 0.1%. The business requires a true positive rate of at least 80%. Which of the following actions is most likely to achieve the required true positive rate while minimizing the increase in false positives?

Accepted Answer

Change the threshold for classifying a transaction as fraud from the default 0.5 to a lower value. (increase hidden layers) may capture more complexity but does not directly increase TPR and could overfit. Option B (SMOTE) rebalances the training set, but the dataset is already balanced, so this is unlikely to improve TPR. Option D (L2 regularization) reduces overfitting but increases bias, which could lower TPR. Option C (change threshold) is the most direct approach: lowering the classification threshold increases the true positive rate, and by tuning, it can achieve 80% TPR with a minimal increase in false positives.

Answer

Increase the number of hidden layers to five to capture more complex patterns

Answer

Use synthetic minority oversampling (SMOTE) to rebalance the training set

Answer

Add L2 regularization to reduce overfitting