AI0-001 Practice Exam 2 — 20 Questions

Question 1

A data scientist notices the model overfits. Which change to the exhibit's configuration would most likely reduce overfitting?

Accepted Answer

Add L2 regularization to dense layers. Adding L2 regularization to dense layers penalizes large weights by adding a squared magnitude term to the loss function, which forces the model to learn simpler patterns and reduces overfitting. This directly addresses the core issue of the model memorizing noise in the training data.

Answer

Remove dropout layers

Answer

Increase learning rate to 0.01

Answer

Increase units in the first dense layer to 512

Question 2

A data scientist notices that a model's performance on the training set is excellent, but validation accuracy is poor. The team used the same dataset for feature engineering and model selection. What is the MOST likely cause?

Accepted Answer

Train/test leakage caused by using the same data for feature engineering and model selection. Reusing the same data for feature engineering and model selection creates data leakage from validation set into training decisions, leading to overfitting and poor generalization.

Answer

The dataset is too small for the model complexity

Answer

The learning rate is too high

Answer

The model is overfitting due to high variance

Question 3

Which TWO of the following are effective techniques for detecting bias in an AI model?

Accepted Answer

Fairness metrics such as equal opportunity difference. Fairness metrics such as equal opportunity difference directly quantify bias by measuring the difference in true positive rates between privileged and unprivileged groups. A value of zero indicates perfect fairness, while non-zero values reveal disparate treatment, making it a standard technique for bias detection in AI models.

Answer

Feature importance scores

Answer

Confusion matrix on the entire dataset

Answer

Cross-validation accuracy

Question 4

A company uses an AI model to screen job applicants. A disparate impact analysis reveals that the model's rejection rate for a protected group is significantly higher than for others. Which THREE actions should the company take to address this?

Accepted Answer

Revisit training data for historical bias and consider reweighting. Revisiting the training data for historical bias and applying reweighting directly addresses the root cause of disparate impact. If the training data contains biased labels or skewed representation of the protected group, the model will learn and amplify those biases. Reweighting adjusts the loss function to give more importance to underrepresented or disadvantaged groups, helping to equalize error rates across groups.

Answer

Ignore the disparity because the model is accurate overall

Answer

Remove all demographic attributes from the dataset

Question 5

A data scientist is building a model to predict credit default using historical loan data. The dataset contains 100,000 records with 50 features, including income, debt-to-income ratio, and loan amount. The target variable is binary (default vs. no default). The goal is to maximize interpretability while maintaining high accuracy. Which algorithm is MOST appropriate?

Accepted Answer

Logistic regression. Logistic regression is interpretable (coefficients show feature impact) and performs well on binary classification with a large dataset. Decision trees are interpretable but may overfit; random forests and gradient boosting are less interpretable.

Answer

Random forest

Answer

Gradient boosting machine

Answer

Decision tree

Question 6

Which TWO deployment strategies allow for testing a new model version before fully rolling it out?

Accepted Answer

Canary deployment. Canary deployment is correct because it routes a small percentage of live traffic to the new model version while the majority continues using the stable version. This allows real-world validation of the new model's performance and error rates under production load before a full rollout, minimizing blast radius if issues arise.

Answer

Shadow deployment

Answer

Direct cutover

Answer

Blue-Green deployment

Question 7

Which THREE factors are most critical to consider when designing a continuous integration/continuous deployment (CI/CD) pipeline for machine learning?

Accepted Answer

Data quality and schema validation. Data quality and schema validation (A) are critical because ML pipelines are highly sensitive to data drift, missing values, and schema mismatches that can silently degrade model performance. Without automated validation at the CI stage, bad data can pass through and corrupt model training or inference, leading to unreliable outputs in production.

Answer

A/B testing framework for comparing models

Answer

Automated unit testing of application code

Question 8

An e-commerce company deploys a recommendation model that must serve predictions with sub-100 ms latency for millions of users during peak hours. The model is a large neural network. Which architecture is most suitable?

Accepted Answer

Use a distributed system with load balancers and model replicas.. Distributing the model across multiple servers with load balancers and replicas allows horizontal scaling to handle millions of concurrent users while maintaining sub-100 ms latency. This architecture provides fault tolerance and can dynamically adjust to peak traffic loads, which is essential for real-time inference with large neural networks.

Answer

Batch process predictions every hour.

Answer

Deploy the model on a single powerful GPU server.

Answer

Use serverless functions with auto-scaling.

Question 9

A data scientist is training a convolutional neural network (CNN) for object detection. The training loss decreases rapidly but then plateaus at a high value, and the validation loss starts increasing. Which action should the scientist take to improve the model?

Accepted Answer

Reduce the model complexity. The training loss decreasing rapidly then plateauing at a high value while validation loss increases is classic overfitting. Reducing model complexity (Option C) directly addresses overfitting by decreasing the number of parameters or applying regularization (e.g., dropout, L2), which forces the network to learn more generalizable features rather than memorizing noise in the training data.

Answer

Increase the learning rate

Answer

Increase the number of epochs

Answer

Add more convolutional layers

Question 10

A financial institution is training a risk assessment model. The dataset includes customer credit scores, income, age, and past loan defaults. During feature engineering, a data engineer creates a new feature 'income_to_debt_ratio'. Which type of feature engineering technique is this?

Accepted Answer

Feature combination. 'income_to_debt_ratio' is created by combining two existing features (income and debt) into a single derived feature. This is a classic example of feature combination (also known as feature crossing or feature construction), where arithmetic operations or logical rules are applied to existing variables to generate new predictive signals. The goal is to capture interactions or relationships that the original features alone may not express linearly.

Answer

Feature encoding

Answer

Feature scaling

Answer

Feature selection

Question 11

A company is choosing between fine-tuning and RAG for a legal document assistant. Which TWO factors would MOST strongly favor RAG over fine-tuning?

Accepted Answer

The legal documents are updated frequently (weekly). RAG allows dynamic retrieval from a changing document base without retraining, and provides citation sources for transparency — critical in regulated domains like law.

Answer

The model needs to understand complex legal terminology

Answer

The queries require deep reasoning across multiple documents

Answer

The company has limited compute budget for training

Question 12

A credit union uses an AI model to approve personal loans. The model was trained on historical data from the past five years. A recent internal review shows that the model approves loans predominantly for white applicants compared to other ethnicities, even when income and credit scores are similar. The credit union wants to comply with fair lending laws without significantly reducing overall approval rates. The data science team has access to the training data. What is the most appropriate remediation step?

Accepted Answer

Resample the training data to ensure balanced representation of ethnicities. Resampling the training data to ensure balanced representation of ethnicities directly addresses the root cause of the bias—skewed historical data—without altering the model's decision logic or approval thresholds. By rebalancing the dataset (e.g., oversampling underrepresented groups or undersampling the majority), the model learns from a more equitable distribution of features, reducing disparate impact while preserving overall approval rates. This approach aligns with fair lending laws by mitigating bias at the data level, which is the most fundamental and effective remediation step.

Answer

Apply a fairness constraint that penalizes the model for disparate impact

Answer

Discontinue the AI model and use manual approval for all loans

Answer

Adjust the approval threshold so that approval rates are equal across ethnic groups

Question 13

A financial institution wants to deploy an AI system that automates loan approvals. Under the EU AI Act, this system would be classified as high-risk. Which of the following is a MANDATORY requirement for high-risk AI systems before market placement?

Accepted Answer

Establishing a risk management system throughout the system lifecycle. The EU AI Act requires a conformity assessment (often including third-party involvement) for high-risk AI systems to ensure compliance with requirements such as data governance, transparency, and human oversight. The other options are either voluntary, not required, or apply to lower-risk tiers.

Answer

Obtaining a CE marking of conformity through self-assessment only

Answer

Acquiring liability insurance for AI-related harms

Answer

Publishing the training dataset in an open repository

Question 14

Which TWO techniques are commonly used to prevent overfitting in deep neural networks?

Accepted Answer

Dropout. Dropout is a regularization technique that randomly drops a fraction of neurons during training, which prevents the network from relying too heavily on any single neuron and forces it to learn more robust features. This reduces overfitting by introducing noise and effectively training an ensemble of sub-networks.

Answer

Using a larger learning rate

Answer

L1 regularization

Answer

Increasing the number of layers

Question 15

A company wants to generate realistic images of new product designs. They have a large dataset of existing product images. Which generative AI approach is MOST suitable for creating novel, high-quality images?

Accepted Answer

Generative adversarial network (GAN). GANs (Generative Adversarial Networks) consist of a generator and discriminator that compete, producing highly realistic images. Diffusion models are also good but GANs are historically the go-to for image generation. VAEs produce blurrier images; LLMs are for text.

Answer

Large language model (LLM)

Answer

Variational autoencoder (VAE)

Answer

Diffusion model

Question 16

A large enterprise is developing an internal LLM-powered assistant that can access the internet and execute code. To mitigate risks from excessive agency (e.g., the model performing unauthorized actions), which THREE security measures should be implemented?

Accepted Answer

Require human-in-the-loop approval for code execution and write operations. Requiring human-in-the-loop approval for code execution and write operations directly enforces a control over the model's agency, preventing it from performing unauthorized actions such as modifying files or executing arbitrary commands. This measure ensures that any action with side effects is vetted by a human operator, mitigating the risk of excessive agency where the LLM could autonomously cause harm.

Answer

Deploy monitoring for anomalous input patterns

Answer

Apply output filtering to block sensitive data in responses

Question 17

Under the GDPR, individuals have the right to not be subject to a decision based solely on automated processing if it produces legal effects. Which of the following is a typical safeguard that organisations must provide to comply with this right?

Accepted Answer

The right to obtain human intervention on the part of the controller. GDPR requires human intervention as a safeguard for automated decisions with legal effects. The right to explanation and human oversight are key. Deleting data is not a safeguard; consent is not sufficient; detailed explanations are part of transparency.

Answer

The right to have all personal data deleted immediately

Answer

The right to receive a detailed mathematical explanation of the model

Answer

The right to demand a more favourable automated decision

Question 18

A team is building a regression model to predict house prices. The dataset includes numerical features (square footage, number of bedrooms) and categorical features (neighborhood, roof type). The categorical features have high cardinality (neighborhood has 200+ unique values). Which encoding strategy should the team use to avoid overfitting and maintain model interpretability?

Accepted Answer

Target encoding with regularization.. Target encoding with regularization is the best choice because it replaces each categorical value with the mean of the target variable for that category, which captures the relationship between the category and house prices. Regularization (e.g., adding a prior or using cross-validation) shrinks the encoded values toward the global mean, preventing overfitting on rare categories (e.g., neighborhoods with only a few houses). This maintains interpretability because each encoded value directly reflects the average price impact of that category, unlike black-box embeddings.

Answer

Label encoding.

Answer

Binary encoding.

Answer

One-hot encoding with feature selection.

Question 19

Which TWO statements correctly describe the difference between supervised and unsupervised learning?

Accepted Answer

Supervised learning requires labeled data. Supervised learning relies on labeled datasets where each training example is paired with an output label, enabling the model to learn a mapping from inputs to outputs. This is a fundamental distinction from unsupervised learning, which works with unlabeled data to find inherent structures or patterns.

Answer

Supervised learning is only used for classification

Answer

Unsupervised learning always requires a target variable

Answer

Supervised learning is a subset of reinforcement learning

Question 20

A machine learning engineer is training a neural network for image classification. The training loss decreases slowly and the model accuracy improves only marginally each epoch. Which hyperparameter adjustment is MOST likely to accelerate convergence?

Accepted Answer

Increase the learning rate. The training loss decreasing slowly and accuracy improving marginally each epoch indicates that the learning rate is too small, causing the optimizer to take very small steps toward the minimum of the loss function. Increasing the learning rate allows the optimizer to take larger steps per update, which accelerates convergence. Option C is correct because adjusting the learning rate directly addresses the step size in gradient descent.

Answer

Add more hidden layers

Answer

Increase the batch size

Answer

Decrease the number of epochs