Machine Learning Fundamentals Assessment — Questions and Answers

Question 1: What is the elbow method used for in clustering?

Evaluating silhouette scores across different algorithms
Identifying outliers before clustering
Selecting the optimal learning rate for k-means
Determining the optimal number of clusters by identifying where inertia reduction slows (Correct answer)

Correct answer: Determining the optimal number of clusters by identifying where inertia reduction slows

The elbow method plots within-cluster inertia against the number of clusters (k) and looks for a kink where adding more clusters yields diminishing returns.

Question 2: Which approach is recommended for troubleshooting feature engineering issues?

Rely solely on past experience
Use systematic isolation and testing methods (Correct answer)
Wait for the problem to resolve itself
Replace all components simultaneously

Correct answer: Use systematic isolation and testing methods

Systematic isolation and testing methodically narrows down the root cause, making troubleshooting efficient and accurate.

Question 3: What is the purpose of the Variance Inflation Factor (VIF) in regression?

Test for heteroscedasticity
Detect multicollinearity among predictors (Correct answer)
Measure prediction error
Evaluate model fit

Correct answer: Detect multicollinearity among predictors

VIF quantifies how much the variance of a coefficient is inflated due to linear correlation with other predictors; VIF > 10 is concerning.

Question 4: In Machine Learning practice, reliability in assessment refers to:

The consistency and reproducibility of results (Correct answer)
The speed of administration
The cost of the assessment tool
The popularity of the instrument

Correct answer: The consistency and reproducibility of results

Reliability refers to the consistency and reproducibility of assessment results when the test is repeated under similar conditions.

Question 5: What is a key characteristic of Naive Bayes classifiers that makes them 'naive'?

They ignore the training labels
They only work with binary classification
They require no hyperparameter tuning
They assume all features are conditionally independent given the class (Correct answer)

Correct answer: They assume all features are conditionally independent given the class

Naive Bayes assumes conditional independence between features given the class label, which simplifies computation but rarely holds in practice.

Question 6: What distinguishes Lasso regression from Ridge regression in terms of feature selection?

Lasso minimizes MSE; Ridge minimizes MAE
Lasso uses L2 penalty; Ridge uses L1
Lasso requires standardization; Ridge does not
Lasso can shrink coefficients exactly to zero; Ridge cannot (Correct answer)

Correct answer: Lasso can shrink coefficients exactly to zero; Ridge cannot

Lasso's L1 penalty produces sparse solutions by setting some coefficients exactly to zero, effectively performing feature selection.

Question 7: What does 'fine-tuning' mean when adapting a pretrained deep learning model?

Pruning unnecessary neurons from the model
Quantizing the model weights for deployment
Continuing to train some or all layers of a pretrained model on a new task (Correct answer)
Training the model from scratch on a new dataset

Correct answer: Continuing to train some or all layers of a pretrained model on a new task

Fine-tuning updates the pretrained weights on new task data, often with a small learning rate, to adapt learned representations without losing general knowledge.

Question 8: What is the main risk of applying target encoding without regularization?

Increased dimensionality
Target leakage causing data leakage (Correct answer)
Underfitting due to information loss
Loss of feature interpretability

Correct answer: Target leakage causing data leakage

Without regularization or cross-fitting, target encoding can leak target information into the training features, causing overfitting.

Question 9: What does a high AUC-ROC score (close to 1.0) indicate about a classification model?

The model has low training accuracy
The model is severely overfitting
The model requires more training data
The model perfectly separates the positive and negative classes (Correct answer)

Correct answer: The model perfectly separates the positive and negative classes

An AUC-ROC near 1.0 means the model ranks positive instances higher than negative ones across all thresholds, indicating excellent discriminative ability.

Question 10: What problem does beam search solve in sequence generation?

It approximates the most probable output sequence by maintaining top-k candidates (Correct answer)
It prevents overfitting during training
It handles padding in batched inputs
It speeds up tokenization

Correct answer: It approximates the most probable output sequence by maintaining top-k candidates

Beam search keeps the top-k (beam width) partial sequences at each step, balancing quality and computational cost versus greedy decoding.

Question 11: What does the term 'perplexity' measure in language model evaluation?

The number of parameters in the model
The diversity of the training vocabulary
The speed of token generation
How well a probability model predicts a sample (Correct answer)

Correct answer: How well a probability model predicts a sample

Perplexity measures how well a language model predicts a held-out text — lower perplexity indicates a better-fitting model.

Question 12: What is 'data leakage' in the context of model evaluation?

Model weights being exposed publicly
Overfitting to the validation set
Training data accidentally deleted
Test set information influencing the training process (Correct answer)

Correct answer: Test set information influencing the training process

Data leakage occurs when information from the test/future data leaks into training, causing overly optimistic evaluation metrics.

Question 13: If a model achieves AUC-ROC of 0.5, what does this indicate?

Perfect classification
Better than random guessing
The model always predicts the negative class
Performance equivalent to random guessing (Correct answer)

Correct answer: Performance equivalent to random guessing

An AUC of 0.5 means the model has no discriminative ability and performs no better than random chance.

Question 14: What is the primary consideration when implementing changes to neural networks?

Speed of implementation
Impact assessment and change management (Correct answer)
Vendor preference
Personal convenience

Correct answer: Impact assessment and change management

Impact assessment and proper change management ensure that modifications do not introduce unexpected problems or service disruptions.

Question 15: Which decision tree splitting criterion measures the average uncertainty weighted by each subset's size?

Gain Ratio
Gini Impurity (Correct answer)
Entropy
Information Gain

Correct answer: Gini Impurity

Gini impurity measures how often a randomly chosen element would be misclassified; weighted average across children is minimized at each split in CART.

Question 16: When comparing two models using cross-validation scores, which statistical test is commonly used?

Paired t-test (Correct answer)
Z-test for proportions
Chi-squared test
ANOVA

Correct answer: Paired t-test

A paired t-test compares cross-validation scores from the same folds across two models, accounting for the paired nature of the data.

Question 17: Which statement about the K-Medoids algorithm is correct compared to K-Means?

K-Medoids is faster than K-Means on large datasets
K-Medoids minimizes the sum of squared distances
K-Medoids cannot handle outliers
K-Medoids centroids must be actual data points (Correct answer)

Correct answer: K-Medoids centroids must be actual data points

In K-Medoids, the representative of each cluster (medoid) must be an actual data point, making it more robust to outliers than K-Means.

Question 18: Which assumption of linear regression states that residuals should not be correlated with each other?

Multicollinearity
Normality of residuals
Homoscedasticity
Independence of errors (Correct answer)

Correct answer: Independence of errors

The independence of errors assumption requires that residuals are uncorrelated — violation often occurs in time-series data.

Question 19: What is the 'bias-variance trade-off' in model evaluation?

The trade-off between precision and recall
The tension between underfitting (high bias) and overfitting (high variance) (Correct answer)
The balance between training speed and model accuracy
The balance between model complexity and interpretability

Correct answer: The tension between underfitting (high bias) and overfitting (high variance)

High bias leads to underfitting while high variance leads to overfitting; the trade-off is finding the model complexity that minimizes both.

Question 20: What is the curse of dimensionality?

The phenomenon where high-dimensional spaces make data sparse, degrading distance-based algorithms (Correct answer)
The inability of models to learn from more than 100 features
The exponential growth of model parameters with depth
The computational cost of training on large datasets

Correct answer: The phenomenon where high-dimensional spaces make data sparse, degrading distance-based algorithms

As dimensionality increases, data becomes increasingly sparse and distances between points become less meaningful, causing many ML algorithms to degrade in performance.

Question 21: In a Support Vector Machine, what is the 'margin'?

The misclassification rate on training data
The number of support vectors divided by total samples
The regularization penalty applied to the weight vector
The distance between the decision boundary and the nearest data points from each class (Correct answer)

Correct answer: The distance between the decision boundary and the nearest data points from each class

The margin is the distance between the hyperplane and the closest data points (support vectors) from each class; SVM maximizes this margin.

Question 22: Which professional attribute is most valued in classification within the Machine Learning field?

Prioritizing personal convenience
Working in isolation
Accountability and commitment to standards (Correct answer)
Avoiding challenging situations

Correct answer: Accountability and commitment to standards

Accountability and commitment to professional standards build trust and ensure consistent, high-quality practice.

Question 23: A classifier achieves 98% accuracy on a dataset where 98% of samples belong to class A. What problem does this illustrate?

Underfitting due to insufficient model complexity
Overfitting to the minority class
Data leakage from test set to training set
The accuracy paradox with imbalanced data (Correct answer)

Correct answer: The accuracy paradox with imbalanced data

A trivial classifier predicting only class A achieves 98% accuracy, demonstrating that accuracy is misleading when class distributions are severely imbalanced.

Question 24: Which initialization strategy for K-Means selects initial centroids that are spread far apart to improve convergence?

K-Means++ (Correct answer)
Random initialization
PCA-based initialization
Forgy method

Correct answer: K-Means++

K-Means++ probabilistically selects each subsequent centroid farther from already-chosen ones, reducing the chance of poor local optima.

Question 25: In k-fold cross-validation with k=5, what percentage of data is used for training in each fold?

95%
20%
50%
80% (Correct answer)

Correct answer: 80%

With k=5, the data is split into 5 folds; each iteration uses 4 folds (80%) for training and 1 fold (20%) for validation.

Question 26: What is the primary advantage of using an ensemble method like Random Forest over a single decision tree?

Faster training time
Simpler model interpretation
Lower variance through averaging multiple trees (Correct answer)
Reduced memory usage

Correct answer: Lower variance through averaging multiple trees

Random Forest reduces variance by averaging predictions from many decorrelated trees built on random subsets of features and data.

Question 27: What is a skip connection (residual connection) in deep learning?

A connection that skips the activation function
A shortcut path that adds a layer's input directly to its output (Correct answer)
A dropout mask applied across multiple layers
A connection that skips regularization

Correct answer: A shortcut path that adds a layer's input directly to its output

Skip connections allow gradients to flow directly through the network, enabling training of very deep networks by alleviating the vanishing gradient problem.

Question 28: Which of the following scenarios will K-means clustering fail to produce satisfactory results? 1) Outliers in the data 2) Data points of various densities 3) Nonconvex data points

2 and 3
1, 2, and 3 (Correct answer)
1 and 2
1 and 3

Correct answer: 1, 2, and 3

Explanations: <br> K-means clustering algorithm fails to give good results when the data contains outliers, the density spread of data points across the data space is different, and the data points follow nonconvex shapes.

Question 29: In Named Entity Recognition (NER), which tagging scheme uses B-, I-, and O- prefixes?

SentencePiece tagging
BIO tagging (Correct answer)
Unigram tagging
BPE tagging

Correct answer: BIO tagging

BIO tagging marks the Beginning of an entity, the Inside (continuation) of an entity, and Outside tokens that are not part of any entity.

Question 30: What is a Variational Autoencoder (VAE) primarily used for compared to a standard autoencoder?

Generating new samples by learning a smooth latent space distribution (Correct answer)
More efficient data compression
Faster training convergence
Better anomaly detection accuracy

Correct answer: Generating new samples by learning a smooth latent space distribution

VAEs impose a probabilistic distribution (typically Gaussian) on the latent space, enabling generation of new, realistic samples by sampling from that distribution.

Question 31: What is transfer learning in the context of deep neural networks?

Converting a model from one framework to another
Transferring data between training and validation sets
Moving a trained model from GPU to CPU
Reusing a pretrained model's learned features as a starting point for a new task (Correct answer)

Correct answer: Reusing a pretrained model's learned features as a starting point for a new task

Transfer learning leverages features learned on large datasets (e.g., ImageNet) and fine-tunes the model on a smaller target dataset, saving time and data.

Question 32: Which imputation strategy is most appropriate for a feature with a heavy right-skewed distribution?

Median imputation (Correct answer)
Mean imputation
Zero imputation
Mode imputation

Correct answer: Median imputation

The median is robust to outliers and skew, making it more representative than the mean for skewed distributions.

Question 33: What does the term 'dead neuron' refer to when using ReLU activations?

A neuron that always outputs zero because its weights push its input permanently negative (Correct answer)
A neuron removed by the dropout process
A neuron whose weights have not been initialized
A neuron that outputs the same value regardless of input

Correct answer: A neuron that always outputs zero because its weights push its input permanently negative

A dead ReLU neuron receives only negative inputs, always outputting zero with zero gradient, meaning it can never recover during training.

Question 34: Which loss function is typically used for training a binary classification model with logistic regression?

Mean Squared Error
Binary Cross-Entropy (Correct answer)
Huber Loss
Hinge Loss

Correct answer: Binary Cross-Entropy

Binary cross-entropy (log loss) measures the divergence between predicted probabilities and true binary labels, making it ideal for logistic regression.

Question 35: In time-series feature engineering, what does a 'lag feature' represent?

The rate of change between consecutive observations
The value of a variable at a previous time step (Correct answer)
The moving average of the series
The Fourier transform of the signal

Correct answer: The value of a variable at a previous time step

A lag feature captures the value of the target or a predictor at an earlier time point, allowing models to learn temporal dependencies.

Question 36: What is the Matthews Correlation Coefficient (MCC) particularly useful for?

Comparing models across different domains
Regression tasks with outliers
Measuring clustering quality
Evaluating classifiers on imbalanced datasets (Correct answer)

Correct answer: Evaluating classifiers on imbalanced datasets

MCC accounts for all four cells of the confusion matrix and provides a balanced measure even when classes are severely imbalanced.

Question 37: What is the role of the minPts parameter in DBSCAN?

It controls the learning rate
It sets the distance threshold for merging
It defines the minimum number of points required to form a dense region (Correct answer)
It sets the maximum number of clusters

Correct answer: It defines the minimum number of points required to form a dense region

A point is a core point in DBSCAN if at least minPts points (including itself) fall within its epsilon-radius neighborhood.

Question 38: Which of the following is NOT a hyperparameter of a Random Forest model?

Learned feature weights of the trees (Correct answer)
Number of trees in the forest
Number of features considered at each split
Maximum depth of each tree

Correct answer: Learned feature weights of the trees

Learned feature weights are parameters determined by the training process, not hyperparameters set before training.

Question 39: What is 'recursive feature elimination' (RFE)?

Removing features with zero variance iteratively
Projecting features onto principal components recursively
Selecting features based on correlation with the target
Training a model and repeatedly removing the least important features until a target count is reached (Correct answer)

Correct answer: Training a model and repeatedly removing the least important features until a target count is reached

RFE fits a model, ranks features by importance, removes the weakest feature(s), and repeats until the desired number of features remains.

Question 40: What is a policy in reinforcement learning?

The transition probabilities between environment states
A mapping from states to actions (or action probabilities) that defines agent behavior (Correct answer)
The reward function provided by the environment
The maximum total reward achievable in an episode

Correct answer: A mapping from states to actions (or action probabilities) that defines agent behavior

A policy defines the agent's behavior by mapping states to actions or probabilities of actions, determining what the agent does in each situation.

Question 41: In the context of NLP, what is a 'zero-shot' setting?

The model is trained on zero examples and evaluated randomly
The model is trained with a learning rate of zero
The model uses zero attention heads
The model solves a task at inference time without any task-specific training examples (Correct answer)

Correct answer: The model solves a task at inference time without any task-specific training examples

Zero-shot evaluation tests a model's ability to perform a new task using only a natural language description, without any labeled examples for that specific task.

Question 42: Which scenario best describes overfitting in a neural network?

Gradients are near zero throughout training
Training loss is low but validation loss is much higher (Correct answer)
Training loss is high and validation loss is high
Training loss decreases while validation loss also decreases steadily

Correct answer: Training loss is low but validation loss is much higher

Overfitting occurs when the model memorizes training data, leading to low training loss but poor generalization on unseen validation data.

Question 43: Which technique helps mitigate catastrophic forgetting when fine-tuning large language models on new tasks?

Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA (Correct answer)
Removing dropout during training
Increasing the learning rate
Using a larger batch size

Correct answer: Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA

PEFT methods like LoRA freeze most pre-trained weights and add small trainable adapter modules, preserving general knowledge while learning new tasks efficiently.

Question 44: Which optimizer adapts the learning rate for each parameter based on past gradients and is commonly used in training transformers?

Adam (Correct answer)
RMSProp
SGD with momentum
Adagrad

Correct answer: Adam

Adam combines momentum and RMSProp, maintaining adaptive per-parameter learning rates using first and second moment estimates of gradients.

Question 45: What is the purpose of a skip connection (residual connection) in ResNets?

To allow gradients and inputs to bypass layers, easing training of very deep networks (Correct answer)
To randomly deactivate entire layers during training
To connect the first layer directly to the output layer
To skip certain training examples deemed too easy

Correct answer: To allow gradients and inputs to bypass layers, easing training of very deep networks

Skip connections add the input of a block directly to its output, providing gradient shortcuts that alleviate vanishing gradients in very deep networks.

Question 46: What is 'feature scaling' and why is it important for algorithms like SVMs and k-NN?

Selecting the most important features; reduces computation
Adding polynomial features to increase model complexity
Removing outliers from the training set
Normalizing feature ranges so no single feature dominates distance calculations (Correct answer)

Correct answer: Normalizing feature ranges so no single feature dominates distance calculations

Feature scaling (e.g., standardization or min-max normalization) ensures all features contribute equally to distance-based calculations, which is critical for SVMs and k-NN.

Question 47: Which of the following best describes 'feature learning' in unsupervised learning?

Automatically discovering useful representations from raw data without labels (Correct answer)
Selecting the most important supervised features
Engineering handcrafted features from domain knowledge
Removing correlated features before classification

Correct answer: Automatically discovering useful representations from raw data without labels

Feature learning (representation learning) involves algorithms like autoencoders and deep belief networks that discover compact, meaningful representations from unlabeled data.

Question 48: Which of the following correctly describes a negative R² value?

The model explains negative variance, which is impossible
The model has negative prediction error
R² can never be negative
The model performs worse than a horizontal mean line (Correct answer)

Correct answer: The model performs worse than a horizontal mean line

Negative R² occurs when the model fits the data worse than simply predicting the mean of the target variable.

Question 49: In Machine Learning practice, what is the primary benefit of using multiple assessment methods?

It is required by law in all cases
It fills more class time
It provides a more complete picture of student understanding (Correct answer)
It eliminates the need for grading

Correct answer: It provides a more complete picture of student understanding

Using multiple assessment methods captures different aspects of student understanding and reduces the limitations of any single measure.

Question 50: In the context of neural networks, what is backpropagation?

The process of feeding input data forward through the network
A technique for initializing network weights
The algorithm that computes gradients of the loss with respect to each weight by applying the chain rule backward through the network (Correct answer)
The method for selecting the optimal learning rate

Correct answer: The algorithm that computes gradients of the loss with respect to each weight by applying the chain rule backward through the network

Backpropagation efficiently computes gradients for all weights by applying the chain rule of calculus backward from the output loss to each layer's parameters.

Question 51: What does mean Average Precision (mAP) measure in object detection evaluation?

Average localization error of bounding box predictions
The mean inference speed across different model sizes
Average pixel accuracy across all semantic classes
The average of per-class Average Precision scores across multiple IoU thresholds and classes (Correct answer)

Correct answer: The average of per-class Average Precision scores across multiple IoU thresholds and classes

mAP aggregates the area under the precision-recall curve for each class and averages across all classes and IoU thresholds, providing a comprehensive detection metric.

Question 52: What problem does ICA (Independent Component Analysis) solve that PCA does not?

Reducing dimensionality while maximizing variance
Separating statistically independent non-Gaussian source signals (Correct answer)
Finding the optimal number of clusters
Removing correlated features

Correct answer: Separating statistically independent non-Gaussian source signals

ICA finds a decomposition into statistically independent components, making it suitable for blind source separation tasks like separating audio signals.

Question 53: We set the gradient to zero to obtain the minimum or maximum of a function because:

Depends on the type of problem
A and B
The value of the gradient at extrema of a function is always zero (Correct answer)
None of these

Correct answer: The value of the gradient at extrema of a function is always zero

Explanation: <br> The gradient of a multivariable function at a maximum point will be the zero vector of the function, which is the single greatest value that the function can achieve.

Question 54: Which of the following best describes 'feature importance' from a Random Forest model?

The correlation coefficient of a feature with the target variable
The average reduction in impurity (e.g., Gini) across all trees when a feature is used to split (Correct answer)
The p-value of a feature in a linear regression
The number of times a feature appears in the training data

Correct answer: The average reduction in impurity (e.g., Gini) across all trees when a feature is used to split

Random Forest feature importance measures how much each feature decreases weighted impurity across all splits in all trees, averaged over the ensemble.

Question 55: What does the 'bias' term represent in the bias-variance tradeoff?

The mean of the training labels
Sensitivity to small fluctuations in training data
Noise in the test dataset
Error due to overly simplistic model assumptions (Correct answer)

Correct answer: Error due to overly simplistic model assumptions

Bias reflects error introduced by assuming a model form that is too simple to capture the true underlying relationship in the data.

Question 56: What problem does an LSTM network solve compared to a vanilla RNN?

It replaces activation functions with attention layers
It removes the need for backpropagation
It handles longer-term dependencies by using gating mechanisms (Correct answer)
It speeds up training with batch normalization

Correct answer: It handles longer-term dependencies by using gating mechanisms

LSTM's cell state and gates (input, forget, output) allow it to retain or discard information over long sequences, mitigating the vanishing gradient problem in RNNs.

Question 57: In support vector regression (SVR), what does the ε (epsilon) parameter control?

The kernel bandwidth for non-linear mappings
The regularization strength of the model
The width of the tube within which errors are ignored (Correct answer)
The learning rate of the optimization

Correct answer: The width of the tube within which errors are ignored

The ε-insensitive tube in SVR means errors smaller than ε incur no penalty, controlling tolerance for deviation from the regression line.

Question 58: In neural architecture search (NAS), what is the goal?

To prune a pretrained model to the smallest size
To automatically discover optimal neural network architectures using search algorithms (Correct answer)
To search for the best learning rate schedule
To manually design the best network topology for a task

Correct answer: To automatically discover optimal neural network architectures using search algorithms

NAS automates the design of neural network architectures by searching over possible configurations using techniques like reinforcement learning or evolutionary algorithms.

Question 59: What is the role of an embedding layer in a neural network processing text?

It converts text to uppercase for normalization
It computes attention scores between word pairs
It applies convolutional filters across word sequences
It maps discrete tokens to dense continuous vector representations (Correct answer)

Correct answer: It maps discrete tokens to dense continuous vector representations

An embedding layer learns a dense vector for each token, capturing semantic relationships in a continuous space.

Question 60: What is the exploding gradient problem and a common mitigation technique?

Weights diverging due to large batches; fixed by smaller learning rates
Gradients growing exponentially; mitigated by gradient clipping (Correct answer)
Gradients becoming zero; solved by ReLU activation
Loss increasing unboundedly; fixed by dropout

Correct answer: Gradients growing exponentially; mitigated by gradient clipping

Exploding gradients occur when backpropagated values grow exponentially; gradient clipping caps gradient norms to keep updates stable.

Machine Learning Fundamentals Assessment

The Machine Learning Fundamentals Assessment covers supervised and unsupervised learning, neural networks, deep learning, feature engineering, model evaluation, regression, classification, clustering, natural language processing, and computer vision techniques used in modern ML applications.

Questions

90 min

Time Limit

70%

To Pass

Checking for saved progress...

Exam Rules

You can skip questions and return to them later
Flag questions for review before submitting
No feedback shown until you submit the entire exam
Unanswered questions count as wrong — answer everything
10 pretest questions are mixed in and don't affect your score
Timer auto-submits when time runs out
Your progress is auto-saved every 30 seconds