Machine Learning (DTSC 3220)

Explanation:

In machine learning, input features are the data provided to the model to make predictions or classifications, such as patient age, blood pressure, or lab results in a healthcare dataset. Model parameters, on the other hand, are the internal variables of the model, like weights and biases, that are learned and adjusted during training to minimize the loss function. The model uses the input features along with these learned parameters to generate predictions. Distinguishing between input features and parameters is essential for understanding how models learn from data.

Correct Answer:

Input features are the data given to the model, whereas model parameters are the internal variables that are learned.

Why Other Options Are Wrong:

Input features are adjustable during training, whereas model parameters are fixed

This is incorrect because input features are fixed data values provided to the model, while model parameters are the ones that are adjusted during training.

Model parameters are manually set before training, whereas input features are learned

This is incorrect because model parameters are learned during training, not manually set. Input features are the raw data provided, not learned variables.

Input features and model parameters are interchangeable terms

This is incorrect because input features and model parameters serve distinct roles; features are the inputs, and parameters are the learnable aspects of the model. They are not interchangeable.

Explanation:

In gradient descent, the learning rate is a hyperparameter that controls the step size for updating the regression coefficients during each iteration. It determines how quickly or slowly the model moves toward the minimum of the cost function. A learning rate that is too large can cause overshooting and divergence, while a learning rate that is too small can lead to very slow convergence. Properly setting the learning rate ensures efficient and stable optimization.

Correct Answer:

It controls the adjustment speed of the regression coefficients towards the minimum of the cost function

Why Other Options Are Wrong:

It specifies the number of iterations to perform

This is incorrect because the learning rate does not determine the number of iterations; it only affects how much the parameters are adjusted in each iteration. The number of iterations is set independently.

It defines the initial values of the regression coefficients

This is incorrect because initial coefficient values are set separately, typically randomly or using heuristics. The learning rate only affects the step size during updates.

It determines the final value of the cost function

This is incorrect because the learning rate influences the path of convergence but does not directly set the final minimum value of the cost function, which is determined by the model and data.

Explanation:

Stochastic Gradient Descent (SGD) updates model parameters by computing the gradient of the objective function for each individual data point and immediately adjusting the parameters. This approach contrasts with batch gradient descent, which calculates the gradient over the entire dataset before updating weights. SGD allows for more frequent updates, which can speed up convergence and introduce beneficial stochasticity that helps the model escape local minima. It is particularly useful for large datasets where computing the full gradient would be computationally expensive.

Correct Answer:

SGD updates parameters by computing the gradient of the objective function for each individual data point and adjusting parameters accordingly.

Why Other Options Are Wrong:

SGD updates parameters by averaging the gradients of the objective function across the entire dataset before making a change

This is incorrect because averaging gradients across the full dataset describes batch gradient descent, not stochastic gradient descent. SGD updates occur after evaluating each single data point rather than the entire dataset.

SGD updates parameters by randomly selecting a subset of data points and using their combined gradient to update the model

This is incorrect because using a subset of data points describes mini-batch gradient descent, which is different from SGD. While both introduce stochasticity, SGD operates on individual data points, not subsets.

SGD updates parameters by fixing the model weights and only adjusting the learning rate based on the performance of the model

This is incorrect because SGD actively updates the model weights using computed gradients. It does not simply adjust the learning rate while leaving weights fixed; the purpose of SGD is to iteratively modify weights to minimize the loss function

Explanation:

In standard supervised learning, labels are typically prepared by a human expert. These labels provide the ground truth outcomes that guide the model during training, allowing it to learn the mapping between input features and target variables. Human experts ensure that the labels are accurate and reliable, which is critical for model performance, particularly in domains like healthcare where precise labeling is essential.

Correct Answer:

Labels are prepared by a human expert.

Why Other Options Are Wrong:

Labels are not required for Supervised learning

This is incorrect because supervised learning fundamentally relies on labeled data to train models. Without labels, the learning task would be unsupervised, not supervised.

Labels are generated by an autonomous agent

This is incorrect because labels generated autonomously without human validation may be inaccurate, and standard supervised learning typically relies on expert-labeled data, especially in high-stakes domains.

Labels can be automatically generated

This is incorrect because although some datasets may have automatically generated labels, standard supervised learning assumes carefully curated and validated labels to ensure accurate model training. Automatic labeling alone is not the default approach.

Explanation:

Unlabeled data is associated with unsupervised machine learning. In unsupervised learning, the dataset does not include target labels or outcomes, and the algorithm must identify patterns, structures, or groupings on its own. Common unsupervised learning tasks include clustering, dimensionality reduction, and anomaly detection. The absence of labels distinguishes unsupervised learning from supervised learning, where labeled data is used to train predictive models.

Correct Answer:

Unsupervised machine learning

Why Other Options Are Wrong:

Unsupervised decisions

This is incorrect because “unsupervised decisions” is not a recognized term in machine learning. The correct terminology is “unsupervised learning,” which refers to algorithms processing unlabeled data.

Unstructured decisions

This is incorrect because unstructured decisions refer to decision-making in business or management contexts, not a type of machine learning algorithm or data association.

Unstructured machine learning

This is incorrect because “unstructured machine learning” is not a standard term. While machine learning can process unstructured data (like text or images), the association with unlabeled data specifically refers to unsupervised learning.

Explanation:

Classification in machine learning is a supervised learning task in which the model is trained to assign input data points to predefined categories or classes based on labeled examples. The model learns patterns from the features of training data to predict the correct class for new, unseen data. Unlike regression, which predicts continuous numerical values, classification deals specifically with discrete outcomes or labels.

Correct Answer:

A type of supervised learning where the goal is to assign input data points to predefined categories or classes.

Why Other Options Are Wrong:

A type of supervised learning where the goal is to predict a continuous target based on input features.

This is incorrect because predicting a continuous target describes regression, not classification. Classification focuses on discrete categories, not numerical values.

A type of reinforcement learning where the goal is to learn an optimal policy for making decisions in an environment.

This is incorrect because reinforcement learning is a separate paradigm in which agents learn through trial and error to maximize cumulative rewards, not to classify data points.

A type of unsupervised learning where the goal is to group similar data points into clusters.

This is incorrect because clustering is an unsupervised learning task. Classification is supervised, requiring labeled data for training.

Explanation:

In softmax regression, the K-dimensional probability vector represents the predicted probabilities for each of the K possible classes for a given instance. Each element of the vector corresponds to the likelihood that the input belongs to a specific class, and all probabilities sum to one. This vector allows the model to make multi-class predictions and is central to calculating the cross-entropy loss during training.

Correct Answer:

It provides the predicted probabilities for each of the K classes associated with an instance.

Why Other Options Are Wrong:

It indicates the total number of features used in the model

This is incorrect because the K-dimensional vector relates to the number of classes, not the number of input features.

It represents the loss function values for each class

This is incorrect because the vector contains predicted probabilities, not loss values. Loss is computed separately, typically using cross-entropy.

It denotes the gradient values for model optimization

This is incorrect because gradients are calculated during backpropagation for optimization purposes. The K-dimensional vector represents output probabilities, not gradient values.

Explanation:

Regression models in machine learning are designed to predict continuous numerical values rather than discrete class labels. For example, they can predict temperatures, prices, or weights. In contrast, classification models predict categorical outcomes, such as classifying emails as spam or not spam. Both regression and classification models typically require labeled data for supervised learning, but regression specifically targets continuous outputs, distinguishing it from classification tasks.

Correct Answer:

Regression models predict continuous numerical values

Why Other Options Are Wrong:

Regression models output categorical labels

This is incorrect because outputting categorical labels is the defining feature of classification models, not regression models.

Regression models require labeled data for training

This is incorrect because, although true, it does not differentiate regression from classification. Both regression and classification are types of supervised learning and require labeled data.

Regression models are only used for binary outcomes

This is incorrect because binary outcomes are a type of classification problem. Regression models predict a continuous range of values, not limited to two outcomes.

Explanation:

The stage of predictive modeling that involves selecting the appropriate algorithm, setting its parameters, and training the model using the training dataset is the Modeling stage. During this phase, data scientists apply machine learning techniques, optimize hyperparameters, and iteratively train models to capture relationships between features and the target variable. This stage is critical for creating an effective predictive model that can generalize well to unseen data

Correct Answer:

Modeling

Why Other Options Are Wrong:

Business understanding

This is incorrect because the business understanding stage focuses on defining objectives, understanding the problem context, and determining project goals. It does not involve algorithm selection or model training.

Deployment

This is incorrect because deployment involves putting the trained model into production to make predictions on new data. It occurs after the model has been trained and evaluated.

Data preparation

This is incorrect because data preparation deals with cleaning, transforming, and organizing data before modeling. While it is essential for successful modeling, it does not include algorithm selection or training.