AI0-001 AI Concepts and Techniques Practice Test 2 — 15 Questions

Question 1

A machine learning engineer is training a convolutional neural network (CNN) for object detection in satellite imagery. The training loss is not decreasing significantly. Which TWO adjustments could help the model converge? (Select TWO)

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Normalize the pixel values to zero mean and unit variance. Option A is correct because normalizing pixel values to zero mean and unit variance (e.g., via standardization or per-channel normalization) conditions the input distribution, which stabilizes and accelerates gradient descent and is a standard prerequisite for CNN convergence. Option E is correct because when training loss plateaus, the optimizer may be overshooting the minimum or stuck in a region where updates are too large; reducing the learning rate (e.g., via a scheduler or by a factor of 2-10) allows finer weight updates and often restores convergence. Option B is not necessarily helpful: dropout is a regularizer that reduces overfitting, and removing it does not address a non-decreasing training loss and may worsen generalization. Option C is incorrect because a smaller batch size mainly reduces memory usage and adds gradient noise; it does not reliably fix convergence and can even destabilize training. Option D is incorrect because increasing the learning rate by 10x typically causes divergence or oscillation rather than convergence when loss is already stagnant.

Answer

Remove dropout layers to allow more gradient flow

Answer

Use a smaller batch size to reduce memory

Answer

Increase the learning rate by 10x

Question 2

A hospital's AI governance committee is reviewing a model that predicts sepsis risk from electronic health records. The vendor states the model was trained with 'unsupervised feature learning on unlabeled vitals, then fine-tuned with a small labeled set.' The committee wants to correctly categorize the training approach used to produce the final predictor. Which learning paradigm BEST describes the complete training pipeline?

Accepted Answer

Semi-supervised learning. Semi-supervised learning is the correct categorization because the pipeline uses abundant unlabeled data for feature learning and a small labeled set for supervised fine-tuning. This hybrid approach is common in healthcare AI where labels are expensive. The other paradigms either lack a labeling component, require reward signals, or involve a distinct source task, none of which fit the described training process.

Answer

Transfer learning

Answer

Reinforcement learning

Answer

Self-supervised learning

Question 3

A company is deploying a large language model for customer support. They want to reduce the number of off-topic or nonsensical responses while maintaining creativity. Which parameter adjustment would BEST achieve this?

Accepted Answer

Decrease temperature to 0.2. Lowering the temperature makes the model's probability distribution sharper, so it favors the highest-probability tokens and produces more deterministic, on-topic output. A value of 0.2 still allows some variation, preserving a degree of creativity while reducing off-topic or nonsensical responses.

Answer

Set top-p to 0.1

Answer

Increase top-k to 100

Answer

Increase temperature to 0.9

Question 4

A startup wants to identify unusual patterns in network traffic to detect potential security breaches. They have a large dataset of normal traffic but very few labeled attacks. Which machine learning approach is MOST suitable?

Accepted Answer

Unsupervised anomaly detection. Unsupervised anomaly detection is the most suitable approach because the startup has a large dataset of normal traffic but very few labeled attacks. This technique learns the baseline of normal behavior from unlabeled data and flags deviations as potential anomalies, which is ideal for detecting unknown or rare attack patterns without requiring labeled attack samples.

Answer

Supervised classification with logistic regression

Answer

Reinforcement learning

Answer

Semi-supervised learning

Question 5

A research team is training a deep learning model for image classification using a small dataset of 1,000 labeled images. They are concerned about overfitting. Which combination of regularisation techniques would be MOST effective?

Accepted Answer

Dropout with a rate of 0.5 and L2 regularisation. Dropout with a rate of 0.5 randomly deactivates half of the neurons during each training step, forcing the network to learn redundant representations and reducing co-adaptation. L2 regularisation adds a penalty proportional to the square of weights to the loss function, discouraging large weights and smoothing the model. Together they combat overfitting from two complementary angles — architectural stochasticity and weight magnitude control — which is the most effective combination among the options.

Answer

Use early stopping without any other regularisation

Answer

L1 regularisation and batch normalisation

Answer

Increase learning rate and use momentum

Question 6

A developer is building a natural language processing system to classify customer reviews as positive, neutral, or negative. They have 50,000 labeled reviews. Which model architecture is MOST appropriate for this task?

Accepted Answer

Fine-tune a pre-trained BERT model. Fine-tuning a pre-trained BERT model is most appropriate because BERT is a transformer-based model pre-trained on a large corpus and can be fine-tuned on the 50,000 labeled reviews to achieve high accuracy with relatively little data. It captures bidirectional context, which is crucial for sentiment classification, and avoids the need for training from scratch.

Answer

Use a convolutional neural network (CNN) on raw text

Answer

Train a recurrent neural network (RNN) from scratch

Answer

Word2vec embeddings followed by logistic regression

Question 7

A machine learning engineer is training a convolutional neural network (CNN) for object detection in satellite imagery. The training loss is not decreasing significantly. Which TWO adjustments could help the model converge? (Select TWO)

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

The learning rate is too low; increase it to 0.01. A learning rate of 0.001 is very low for many logistic regression implementations, causing the gradient descent algorithm to take extremely small steps toward the minimum of the loss function. This results in a slow decrease in loss because each weight update is minimal. Increasing the learning rate to 0.01 allows larger steps per iteration, accelerating convergence without typically causing divergence in well-scaled data.

Answer

The learning rate is too high; decrease it to 0.0001

Answer

The model is overfitting; add L2 regularisation

Answer

The batch size is too large; reduce it

Question 8

A team is training a generative adversarial network (GAN) to generate realistic images of furniture. The generator loss decreases sharply while the discriminator loss increases. What is the MOST likely issue and recommended action?

Accepted Answer

The generator is too strong; train the discriminator more frequently. When the generator loss decreases sharply while the discriminator loss increases, the generator is producing samples realistic enough to fool the discriminator consistently. This indicates the discriminator is not learning effectively, so the recommended action is to strengthen the discriminator by training it more frequently (or increasing its capacity). Training the discriminator more frequently gives it more opportunities to distinguish real from fake, restoring balance in the adversarial game.

Answer

Mode collapse has occurred; increase the generator's learning rate

Answer

The discriminator is overfitting; decrease its capacity

Answer

The learning rates are too high; reduce both

Question 9

A data scientist is selecting a model for a binary classification task where interpretability is critical because of regulatory requirements. The dataset has 20 features and 10,000 samples. Which model is MOST appropriate?

Accepted Answer

Decision tree. A decision tree is inherently interpretable: its if-then-else splits can be visualized and explained to regulators, auditors, or customers. With only 20 features and 10,000 samples, a single tree is also computationally adequate and unlikely to overfit catastrophically if pruned. This makes it the best fit when interpretability is a hard requirement.

Answer

Neural network (MLP)

Answer

Gradient boosting machine

Answer

Random forest classifier

Question 10

A developer is using a pre-trained BERT model for a question-answering system. They want to ensure the model can handle out-of-vocabulary words. Which component of the BERT architecture is responsible for this?

Accepted Answer

WordPiece tokenisation. WordPiece tokenisation is the component of BERT that handles out-of-vocabulary (OOV) words by breaking them into subword units (e.g., 'playing' → 'play' + '##ing'). This allows the model to represent any word, even unseen ones, as a sequence of known subword tokens, ensuring no word is truly out of vocabulary.

Answer

Positional encoding

Answer

Feed-forward layers

Answer

Attention mechanism

Question 11

A team is training a recurrent neural network (RNN) with LSTM units to predict stock prices. The validation loss is significantly higher than the training loss. Which action is MOST likely to reduce the gap?

Accepted Answer

Increase the dropout rate in LSTM layers. A significant gap between training and validation loss indicates overfitting, where the model memorizes training data but fails to generalize. Increasing dropout in LSTM layers is a regularization technique that randomly deactivates neurons during training, forcing the network to learn more robust features and reducing overfitting. This directly addresses the gap by improving validation performance.

Answer

Increase the number of LSTM units

Answer

Increase the number of training epochs

Answer

Reduce the sequence length

Question 12

A company is deploying a chatbot using a large language model. They want to mitigate the risk of prompt injection attacks. Which TWO measures should be implemented?

Accepted Answer

Implement input validation and sanitisation. Option A is correct because input validation and sanitisation directly strip or neutralise adversarial payloads (e.g., embedded instructions, delimiter-breaking characters, or encoded jailbreak strings) before they reach the LLM, reducing the attack surface for prompt injection. Option B is correct because a strict system prompt establishes immutable behavioural boundaries and instruction hierarchy, making it harder for user-supplied text to override the model's intended role or exfiltrate system instructions. Option C is not a reliable mitigation because fine-tuning on safe examples does not prevent novel injection payloads at inference time and can even be undone by adversarial prompting. Option D is irrelevant, since a larger context window merely allows more tokens to be processed and does not filter or constrain malicious instructions. Option E only caps response size and does nothing to stop an injected prompt from altering the model's behaviour or leaking data.

Answer

Fine-tune the model on safe conversational examples

Answer

Use a larger context window

Answer

Limit the maximum output token length

Question 13

A data scientist is evaluating a binary classifier for a medical diagnosis task. The dataset is imbalanced with 5% positive cases. Which THREE metrics should the data scientist consider for a comprehensive evaluation?

Accepted Answer

Precision. Precision (A) is correct because it measures the proportion of predicted positives that are truly positive, which is critical in an imbalanced medical diagnosis task where false positives carry real cost. Recall (E) is correct because it measures the proportion of actual positives that are correctly identified, ensuring the classifier does not miss the rare 5% positive cases. F1 score (B) is correct because it is the harmonic mean of precision and recall, providing a single balanced metric that is far more informative than accuracy on skewed class distributions. Accuracy (C) is not appropriate here because a trivial model predicting all negatives would achieve 95% accuracy while detecting zero positive cases. Perplexity (D) does not belong because it is a language-model metric measuring how well a probability distribution predicts a sample, not a classification performance measure.

Answer

Accuracy

Answer

Perplexity

Question 14

A company wants to build a system that can generate new product images for an online catalog. Which TWO generative AI approaches are most suitable?

Accepted Answer

Diffusion models. Diffusion models (A) are correct because they generate high-fidelity, photorealistic images by iteratively denoising random noise, making them ideal for producing new product images for an online catalog. Generative Adversarial Networks (C) are also correct because a generator-discriminator pair can synthesize realistic product imagery and can be trained to match a catalog's visual style. Variational autoencoders (B) are not the best fit here because their outputs tend to be blurrier and less photorealistic than diffusion or GAN results. BERT-based models (D) are encoder-only language models for understanding text, not image generation. GPT-style language models (E) generate text, not images, so they do not meet the requirement.

Answer

Variational autoencoders (VAEs)

Answer

BERT-based model

Answer

GPT-style language model

Question 15

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

Accepted Answer

Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store. Retrieval-Augmented Generation (RAG) is the most appropriate approach because it allows the chatbot to answer questions based on the latest policy documents without retraining the model. By indexing the documents in a vector store and retrieving relevant chunks at query time, RAG provides up-to-date, contextually accurate answers while keeping the underlying LLM static, which avoids the cost and complexity of monthly retraining.

Answer

Use a larger foundation model with a longer context window and paste all documents into each prompt

Answer

Fine-tune a base LLM on the policy documents monthly

Answer

Train a custom model from scratch on the policy documents each month