Machine Learning (DTSC 3220)

Explanation:

The key difference between supervised and unsupervised learning lies in the presence of labeled data and the task being performed. Supervised learning uses labeled data to predict an output value based on input features. The model learns the relationship between inputs and known outputs to make accurate predictions on new data. In contrast, unsupervised learning works with unlabeled data and focuses on analyzing and discovering patterns, structures, or relationships within the dataset, such as clustering similar data points or identifying underlying data distributions.

Correct Answer:

Supervised learning is used to predict an output value, while unsupervised learning is used to analyze and learn patterns in that data

Why Other Options Are Wrong:

Supervised learning is used to analyze patterns in data, while unsupervised learning is used to predict an output value

This is incorrect because it reverses the roles of supervised and unsupervised learning. Supervised learning predicts outputs, while unsupervised learning finds patterns.

Supervised learning is used to measure input variables, while unsupervised learning is used to measure output variables

This is incorrect because both approaches work with input variables; the distinction is based on the presence of labeled outputs, not on measuring input versus output.

Supervised learning is used to measure output variables, while unsupervised learning is used to measure input variables

This is incorrect because unsupervised learning does not “measure” input variables; it identifies patterns and structures in the data without labels.

None of the above

This is incorrect because one of the provided options correctly defines the difference between supervised and unsupervised learning.

Explanation:

The primary advantage of using the log likelihood in MLE is that it simplifies the optimization process by converting products into sums. The likelihood function often involves the product of many probabilities, which can be computationally challenging and prone to numerical underflow. By taking the natural logarithm, the product of probabilities becomes a sum of log probabilities, making differentiation and optimization more straightforward. This allows for easier calculation of parameter estimates that maximize the likelihood function.

Correct Answer:

It simplifies the optimization process by converting products into sums.

Why Other Options Are Wrong:

It allows for the direct computation of probabilities without transformation

This is incorrect because the log likelihood does not compute probabilities directly; it transforms the likelihood for easier optimization. Probabilities still need to be derived from the model parameters.

It provides a more complex representation of the data

This is incorrect because taking the log actually simplifies the representation mathematically, rather than making it more complex. The purpose is to make optimization manageable, not to complicate the data.

It eliminates the need for parameter estimation

This is incorrect because MLE is inherently about estimating parameters. The log likelihood does not remove the need for parameter estimation; it only facilitates the process of finding the parameters that maximize the likelihood.

Explanation:

Supervised learning is a machine learning approach where models are trained using labeled data, meaning each input has a corresponding output or target. The model learns the mapping between inputs and outputs so that it can predict outcomes for new, unseen data. This contrasts with unsupervised learning, which does not use labels and focuses on discovering patterns, structures, or groupings within the data. Supervised learning is fundamental for tasks like regression and classification.

Correct Answer:

A method that uses labeled data to train models for predicting outcomes.

Why Other Options Are Wrong:

A technique that identifies patterns in data without any labels.

This is incorrect because it describes unsupervised learning, not supervised learning. Supervised learning relies on labeled data to guide the model’s predictions.

An approach that focuses solely on clustering similar data points.

This is incorrect because clustering is an unsupervised learning technique. Supervised learning aims to predict outputs rather than group similar data points.

A process that requires no prior knowledge of the data structure.

This is incorrect because supervised learning relies on labeled data to understand the relationship between inputs and outputs. The process does require structured data with known targets.

Explanation:

In gradient descent, updates to the model parameters continue iteratively until a stopping criterion has been met. Stopping criteria can include reaching a maximum number of iterations, achieving a sufficiently small change in the loss function between iterations, or when the gradient magnitude falls below a predefined threshold. These criteria ensure that the algorithm stops when further updates no longer meaningfully improve the model, rather than continuing indefinitely or stopping prematurely.

Correct Answer:

When a stopping criterion has been met

Why Other Options Are Wrong:

When the gradient's direction flips

This is incorrect because the direction of the gradient can change during optimization without indicating convergence. Gradient descent relies on following the gradient to minimize the loss, so occasional direction changes are normal.

Wen the model has arrived in a local minimum of the loss function

This is incorrect because the algorithm may stop before or after reaching a local minimum depending on the stopping criteria. While a local minimum may be reached, gradient descent does not explicitly stop only at local minima.

When the model fits all points

This is incorrect because perfectly fitting all points is often impossible and unnecessary, especially in cases of noisy data. Overfitting can occur if the model tries to fit every data point exactly, which is not the goal of gradient descent.

Explanation:

Classification tasks aim to assign categorical labels to input data based on learned patterns from labeled training data. The model learns from examples where the correct category is known and predicts the class for new, unseen instances. This distinguishes classification from regression, which predicts continuous outputs, and from clustering or dimensionality reduction, which focus on unsupervised analysis or data representation rather than categorical prediction.

Correct Answer:

To assign categorical labels to data instances

Why Other Options Are Wrong:

To predict continuous outcomes based on input features

This is incorrect because predicting continuous outputs is the goal of regression tasks, not classification.

To group similar data points into clusters

This is incorrect because clustering is an unsupervised learning task, not a classification task. Classification requires labeled outcomes.

To reduce dimensionality of the dataset

This is incorrect because dimensionality reduction focuses on simplifying data representation, not on predicting class labels for instances.

Explanation:

In the context of Generalized Linear Models (GLMs), logistic regression is specifically designed for predicting binary outcomes, where the target variable has two possible classes (e.g., yes/no, 0/1). It models the probability of one class occurring as a function of the predictor variables using the logistic (sigmoid) function. This allows the output to be interpreted as a probability, which can then be used for classification. Logistic regression is not suitable for continuous outputs, variance modeling, or predicting exact counts.

Correct Answer:

Predicting binary outcomes.

Why Other Options Are Wrong:

Predicting continuous outcomes.

This is incorrect because continuous outcomes are predicted using linear regression or other regression models, not logistic regression, which is designed for categorical binary targets.

Modeling variance with the mean of the predictor.

This is incorrect because while GLMs can model variance structures, logistic regression focuses on modeling probabilities of binary outcomes, not variance-mean relationships.

Predicting the exact count of an event.

This is incorrect because predicting counts is typically handled by Poisson regression, not logistic regression. Logistic regression predicts probabilities of classes, not exact event counts.

Explanation:

In supervised machine learning, the distinction between classification and regression lies in the type of output variable. Classification models are designed to assign data points to discrete categories or classes, such as predicting whether a patient has a disease or not. Regression models, on the other hand, predict continuous numerical values, such as forecasting blood pressure levels or predicting house prices. Understanding this distinction is fundamental for selecting the appropriate modeling approach based on the nature of the target variable.

Correct Answer:

Classification assigns data points to categories; regression predicts continuous values.

Why Other Options Are Wrong:

Classification predicts continuous values; regression assigns data points to categories.

This is incorrect because it reverses the definitions of classification and regression. Classification predicts categories, not continuous values.

Classification and regression both predict continuous values.

This is incorrect because only regression predicts continuous outputs. Classification predicts discrete class labels, not continuous values.

Classification and regression both assign data points to categories.

This is incorrect because regression does not assign categories. Regression outputs numerical predictions, making this statement inaccurate.

Explanation:

The objective function, also known as the loss function, is a critical component in machine learning optimization. It quantifies how well the model’s predictions align with the actual target values. During training, the model parameters are adjusted to minimize (or maximize, depending on context) this function. By measuring the model’s performance, the objective function guides the optimization process, ensuring that the model improves accuracy and generalization over time.

Correct Answer:

It measures the performance of the model

Why Other Options Are Wrong:

It defines the model architecture

This is incorrect because the model architecture determines the structure of the neural network or algorithm, not the objective function. The objective function evaluates performance rather than designing the model.

It selects the features for training

This is incorrect because feature selection is a separate process from the objective function. The objective function does not determine which inputs to use; it evaluates model predictions based on given features.

It determines the training dataset size

This is incorrect because the objective function does not control how much data is used for training. Dataset size is an independent decision related to data preparation, not performance measurement.

Explanation:

Predicting the price of a stock 6 months from now is not a classification problem; it is a regression problem because the target variable is continuous. Classification problems involve predicting discrete categories or classes, such as whether a medical condition is present, which DNA sequence is associated with a disease, or whether a transaction is fraudulent. Regression problems, in contrast, involve estimating numerical values.

Correct Answer:

A financial institution wants to predict the price of a stock 6 months from now, on the basis of company performance measures and economic data.

Why Other Options Are Wrong:

A person arrives at the emergency room with a set of symptoms and a medical condition has to be attributed

This is incorrect because attributing a medical condition to a patient is a classification problem. The model predicts discrete categories, such as specific diseases.

Bioinformatics researchers want to figure out which DNA sequences is associated with a certain disease

This is incorrect because this is a classification task. The DNA sequences are classified as being associated with the disease or not.

An online banking service has to decide whether or not a transaction being performed is fraudulent

This is incorrect because determining if a transaction is fraudulent involves predicting one of two categories: fraudulent or non-fraudulent, which is a binary classification problem.