Pattern recognition: a primer

• A. Storing large amounts of data without organizing it
• B. Identifying regularities, structures, or repeated relationships in data ✓
• C. Encrypting information so that it cannot be read by others
• D. Randomly guessing outcomes based on past events

Why: Pattern recognition is the process of detecting regularities, structures, or recurring relationships within data, which allows us to classify, predict, or make sense of new inputs.

• A. Calculating the square root of a number
• B. Recognizing a friend's face in a crowd ✓
• C. Writing a computer program from scratch
• D. Measuring the temperature of a room

Why: Recognizing a friend's face involves matching visual features against stored patterns learned from experience, which is a classic example of pattern recognition.

• A. A bug in the algorithm that causes errors
• B. An individual measurable property of the data used for analysis ✓
• C. The final output produced by a model
• D. The hardware used to run the computations

Why: A feature is an individual measurable property or characteristic of the data, such as the height of a person or the brightness of a pixel, that a model uses to identify patterns.

• A. Supervised learning uses labeled examples, while unsupervised learning finds structure in unlabeled data ✓
• B. Supervised learning requires no data, while unsupervised learning requires large datasets
• C. Supervised learning is only used for images, while unsupervised learning is only used for text
• D. There is no difference; the terms are interchangeable

Why: Supervised learning trains on examples paired with correct labels, whereas unsupervised learning discovers patterns or groupings in data without any provided labels.

• A. It allows systems to generalize what they learned from past examples to new situations ✓
• B. It ensures that new data is always identical to old data
• C. It eliminates the need for any data collection
• D. It guarantees perfect accuracy on every input

Why: Good pattern recognition captures underlying regularities rather than memorizing specific examples, so it can apply what was learned to new inputs it has never encountered before.

• A. Similarity
• B. Proximity ✓
• C. Closure
• D. Continuity

Why: The Gestalt principle of proximity states that objects positioned near each other are perceived as a single group, even when other features differ.

• A. Pareidolia ✓
• B. Habituation
• C. Binocular rivalry
• D. Figure-ground reversal

Why: Pareidolia is the tendency to perceive meaningful patterns, especially faces, in random or ambiguous visual stimuli.

• A. Closure ✓
• B. Similarity
• C. Common fate
• D. Symmetry

Why: Closure is the Gestalt principle by which the mind completes incomplete shapes, perceiving them as whole figures.

• A. Figure-ground relationship ✓
• B. Visual acuity
• C. Contrast adaptation
• D. Saccadic masking

Why: Figure-ground perception describes how the visual system separates a shape (figure) from its surroundings (ground), which in ambiguous images can flip.

• A. Their name is always spoken louder than other words
• B. Familiar and personally significant stimuli capture attention more readily ✓
• C. The brain stops processing all sounds except speech
• D. Names activate the visual cortex directly

Why: Personally meaningful stimuli, such as one's own name, receive preferential attentional processing, which is why they can break through a noisy environment.

• A. Median
• B. Mode
• C. Mean ✓
• D. Range

Why: The mean is calculated by summing all values, so a single extreme value pulls the average toward it. The median and mode depend on position or frequency rather than magnitude, so they are far more resistant to outliers.

• A. The dataset contains only negative values
• B. The dataset has no spread because all values are identical ✓
• C. The mean is equal to the median
• D. The dataset contains no outliers

Why: Standard deviation measures how far values spread from the mean. A value of 0 means every observation equals the mean, so there is no variation at all. The other options do not follow from this fact.

• A. To make the dataset smaller and faster to process
• B. To evaluate how well a model generalizes to data it has not seen ✓
• C. To ensure the training data has more outliers
• D. To guarantee the model will achieve 100% accuracy

Why: A model evaluated only on the data it was trained on can simply memorize it. A held-out test set estimates performance on new data, which is the real goal of pattern recognition.

• A. 68%
• B. 95% ✓
• C. 99.7%
• D. 50%

Why: Values 40 and 60 are exactly two standard deviations below and above the mean (50 minus 10 and 50 plus 10). Under a normal distribution, about 95% of values lie within two standard deviations of the mean.

• A. A model performs poorly on both training and test data
• B. A model learns noise and specific details of the training data, hurting performance on new data ✓
• C. A model uses too few features to capture the underlying pattern
• D. A model trains faster than expected because the dataset is small

Why: Overfitting occurs when a model captures random noise in the training data as if it were a real pattern. It performs well on training data but poorly on unseen data. Using too few features describes underfitting, not overfitting.

• A. To collect as much raw data as possible before training a model
• B. To transform raw input into informative, compact representations that highlight relevant characteristics ✓
• C. To remove all noise from the dataset by deleting samples
• D. To assign class labels to each input example

Why: Feature extraction converts raw data (such as pixels or audio samples) into representations that capture the properties useful for distinguishing patterns, while reducing redundancy and dimensionality.

• A. The output probability of a neural network's final layer
• B. Edge orientation histograms computed from image gradients ✓
• C. The randomly initialized weights of a classifier
• D. The total number of images in the dataset

Why: Edge orientation histograms (as used in HOG descriptors) are classic hand-crafted features designed by engineers, computed directly from pixel gradients, unlike learned or model-internal quantities.

• A. The feature never changes its numerical value for any input
• B. The feature stays the same under certain transformations, such as translation or rotation ✓
• C. The feature requires no normalization before use
• D. The feature is calculated identically across all classifiers

Why: Invariance means a representation remains unchanged when the input undergoes specific transformations, such as shifting or rotating an object, so recognition does not depend on those variations.

• A. It increases the number of features in the dataset
• B. It prevents features with larger numeric ranges from dominating distance-based or gradient-based computations ✓
• C. It guarantees that the classifier will achieve 100% accuracy
• D. It converts all features into categorical labels

Why: Features measured on different scales can bias algorithms that rely on distances or gradients. Normalization puts features on comparable scales so each contributes fairly.

• A. To increase the number of features so the model sees more detail
• B. To remove redundant or less informative dimensions while preserving important structure ✓
• C. To convert numerical data into text strings
• D. To ensure every data point has exactly the same label

Why: Dimensionality reduction, such as PCA, compresses data into fewer dimensions while keeping the variation most useful for recognition, which lowers computational cost and can reduce overfitting.

• A. The maximum number of training examples a model can handle
• B. The surface in feature space that separates regions assigned to different classes ✓
• C. The minimum accuracy a classifier must achieve to be useful
• D. The learning rate used during gradient descent

Why: A decision boundary is the set of points in feature space where the classifier is indifferent between classes; it partitions the space into regions, each associated with one predicted class.

• A. A curve that wraps around clusters of data points
• B. A straight line in two dimensions or a hyperplane in higher dimensions ✓
• C. A boundary that changes its shape for every new data point
• D. A boundary defined only by the mean of each class

Why: A linear decision boundary is a straight line (2D) or hyperplane (higher dimensions) given by an equation of the form w·x + b = 0, which separates the feature space into two half-spaces.

• A. Where the predicted probability equals 0.5 ✓
• B. Where the predicted probability equals 1
• C. Where the feature values are all zero
• D. Where the training loss is minimized

Why: The decision boundary is the set of points where the predicted probability equals the threshold, here 0.5. For logistic regression, this corresponds to w·x + b = 0, a linear boundary.

• A. The model underfits and performs poorly on training data
• B. The model overfits, fitting noise and generalizing poorly to new data ✓
• C. The decision boundary becomes a straight line automatically
• D. The number of classes is reduced to two

Why: An overly complex boundary can wrap around individual training points, including noise, leading to overfitting: strong training performance but weak performance on unseen data.

• A. A single straight line
• B. A nonlinear boundary, such as a circle ✓
• C. A boundary parallel to the x-axis only
• D. No boundary is needed because the classes overlap

Why: Concentric classes cannot be separated by any straight line. A nonlinear boundary such as a circle (or a nonlinear model or kernel transformation) is required to separate the inner and outer rings.

• A. To predict a labeled target variable from input features
• B. To group similar data points together without using labels ✓
• C. To reduce the dimensionality of labeled training data
• D. To measure the accuracy of a trained classifier

Why: Clustering seeks to discover natural groupings in data based on similarity, without any predefined labels guiding the process.

• A. It builds a tree by recursively splitting data on feature thresholds
• B. It assigns each point to the nearest of k centroids and iteratively updates the centroids ✓
• C. It estimates the probability density of data using a mixture of Gaussians
• D. It connects points into a graph and removes the longest edges

Why: K-means alternates between assigning points to their nearest centroid and recomputing each centroid as the mean of its assigned points, until assignments stabilize.

• A. The true class labels of each data point
• B. The number of clusters k ✓
• C. The learning rate for gradient descent
• D. The number of hidden layers in the network

Why: K-means requires the number of clusters k as an input parameter; the algorithm does not determine it automatically.

• A. The best distance metric to use
• B. A reasonable number of clusters k ✓
• C. Which features are most correlated
• D. Whether the data contains missing values

Why: The elbow method plots within-cluster sum of squares against k and looks for the point where adding more clusters yields diminishing improvement.

• A. It always requires the number of clusters to be specified in advance
• B. It produces a tree-like structure (dendrogram) showing nested groupings at multiple levels ✓
• C. It can only be applied to numerical data with fewer than ten features
• D. It assigns each point to exactly one centroid that is never updated

Why: Hierarchical clustering builds a dendrogram representing how clusters merge or split across levels, and the number of clusters can be chosen afterward by cutting the tree.

• A. Precision
• B. Recall ✓
• C. Specificity
• D. Accuracy

Why: Recall (also called sensitivity or true positive rate) is the fraction of actual positives correctly identified: TP / (TP + FN). Precision instead measures how many predicted positives are truly positive.

• A. 0.60
• B. 0.75 ✓
• C. 0.80
• D. 0.20

Why: Precision = true positives / all predicted positives = 60 / 80 = 0.75.

• A. Accuracy cannot be calculated for imbalanced data.
• B. A model that always predicts the majority class achieves 95% accuracy without learning any useful pattern. ✓
• C. Accuracy always decreases as the number of classes increases.
• D. Accuracy only applies to regression problems.

Why: With a 95% majority class, a trivial classifier that always predicts that class scores 95% accuracy while failing to detect the minority class at all. Metrics such as precision, recall, or F1 give a clearer picture.

• A. The exact number of parameters in the model
• B. How well the model is likely to generalize to unseen data ✓
• C. The training time of the algorithm
• D. The size of the original dataset

Why: Cross-validation splits data into k parts, training on k-1 and testing on the remaining part in rotation. Averaging the results gives a more reliable estimate of generalization performance than a single train-test split.

• A. The model performs poorly on both training and test data.
• B. The model performs well on training data but significantly worse on new, unseen data. ✓
• C. The model takes too long to train on a standard computer.
• D. The model uses too few input features to make any predictions.

Why: Overfitting occurs when a model learns noise and specific details of the training set, so its training performance is high while its performance on unseen data drops noticeably.