Matthew Sutherland 5/24/25 Matthew Sutherland 5/24/25

AI Practitioner Glossary

A retro-futuristic sci-fi control room powered by analog machines in a dystopian yet utopian fusion style. The room is filled with glowing cathode ray tube monitors, flickering scanner displays, oscilloscopes, and vacuum tube computers. The atmosphere is dim, with deep shadows and hazy light from orange and green neon lights. A central terminal projects a glitchy hologram of a digital glossary entry, partially corrupted. Surveillance drones hover near the ceiling, their red lenses blinking. The walls are covered in peeling posters of utopian AI propaganda. The scene combines analog textures with digital decay, evoking a noir surveillance vibe in a forgotten AI facility.

48 Evaluation Metric

A standard to assess model performance.

Real-world use:

Used to compare regression models via RMSE, classification models via F1-score, or ranking systems via MAP scores.

56 F-measure (F1 Score)

The harmonic mean of precision and recall.

Real-world use:

Used in binary classification evaluation when you need to balance precision and recall, especially in imbalanced datasets like fraud detection.

Common mistake:

Using F1-score on highly imbalanced datasets where other metrics like AUC-ROC might be more appropriate.

61 GAN (Generative Adversarial Network)

A neural network framework with two models — generator and discriminator — competing to improve outputs.

Real-world use:

Used to generate realistic human faces from random noise, create synthetic training data, and develop deepfake technology for entertainment.

69 Hidden Layer

A layer in a neural network between input and output layers.

Real-world use:

Used to extract abstract features from input data, enabling deep networks to learn complex patterns in image recognition and natural language processing.

62 Generalization

A model's ability to perform well on unseen data.

Real-world use:

Used to evaluate real-world effectiveness of trained models, ensuring they work on new customers, images, or situations not seen during training.

Common mistake:

Confusing training performance with generalization ability, leading to overconfident models that fail in production.

66 Ground Truth

The actual labels used to compare model predictions against.

Real-world use:

Used to evaluate accuracy in image classification tasks, comparing model predictions against expert-verified diagnoses in medical imaging.

73 Imbalanced Dataset

A dataset where some classes are over- or under-represented.

Real-world use:

Used to describe fraud detection data with few fraud cases (1%) versus legitimate transactions (99%), requiring special handling techniques.

Common mistake:

Using accuracy as the primary metric for imbalanced datasets, which can be misleading due to class distribution.

67 Hashing Trick

A technique to convert categorical variables into fixed-length vectors.

Real-world use:

Used in scalable text classification where vocabulary size is huge, mapping words to fixed-size feature vectors for efficient processing.

68 Heuristic

A problem-solving approach using practical methods, not guaranteed to be optimal.

Real-world use:

Used in AI for game strategy estimation, route planning algorithms, and quick decision-making when optimal solutions are computationally expensive.

74 Imputation

Filling in missing data values.

Real-world use:

Used to handle nulls in medical record datasets by replacing missing values with statistical estimates like mean, median, or predicted values.

Common mistake:

Using simple mean imputation for all missing data without considering the underlying patterns or reasons for missingness.

75 Inductive Learning

Learning from labeled data to generalize for new data.

Real-world use:

Used in standard supervised machine learning, training models on historical sales data to predict future sales patterns.

76 Information Gain

A metric to decide feature splits in decision trees.

Real-world use:

Used in building classification trees, selecting which feature to split on based on how much it reduces uncertainty in the target variable.

77 Instance-Based Learning

Learning that memorizes training instances rather than generalizing.

Real-world use:

Used in k-nearest neighbors (KNN) for recommendation systems, finding similar users or products based on stored historical data.

78 Interquartile Range (IQR)

A measure of statistical dispersion between the 25th and 75th percentiles.

Real-world use:

Used in detecting outliers in data by identifying values that fall outside 1.5 × IQR from the quartiles, common in financial data analysis.

Core Algorithms

85 k-Means

A clustering algorithm that partitions data into k clusters.

Real-world use:

Used in market segmentation to group customers based on purchasing behavior, demographics, and preferences for targeted marketing campaigns.

Common mistake:

Choosing k arbitrarily without using methods like the elbow method to determine optimal cluster number.

86 k-Nearest Neighbors (KNN)

A classification method based on closest training examples.

Real-world use:

Used in handwriting recognition, recommendation systems, and image classification by finding the most similar training examples.

94 Linear Regression

A method to model the relationship between variables with a straight line.

Real-world use:

Used to predict housing prices based on features like square footage, location, and number of bedrooms using a linear relationship.

95 Logistic Regression

A model for binary classification problems.

Real-world use:

Used to predict if a customer will churn, whether an email is spam, or if a patient has a disease based on input features.

97 LSTM (Long Short-Term Memory)

A type of RNN good at remembering long sequences.

Real-world use:

Used in speech recognition and time series forecasting, handling sequences where long-term dependencies matter, like stock price prediction.

93 LeNet

One of the earliest convolutional neural networks.

Real-world use:

Used in digit recognition tasks, pioneering the application of CNNs for handwritten digit classification in postal services.

83 JSON (JavaScript Object Notation)

A lightweight data format used for data exchange.

Real-world use:

Used to structure data in REST APIs for ML apps, enabling seamless data transfer between web services and machine learning models.

84 k-Fold Cross-Validation

A technique that divides data into k parts and trains/testing k times.

Real-world use:

Used in robust model evaluation to ensure models generalize well, typically using 5-fold or 10-fold validation in machine learning competitions.

87 Kernel Function

A function used in SVM to enable non-linear classification.

Real-world use:

Used in separating data that isn't linearly separable, like classifying images or text where complex decision boundaries are needed.

89 Label

The ground-truth outcome associated with a data point.

Real-world use:

Used in supervised learning for prediction targets, like "spam" or "not spam" for emails, or house prices for real estate data.

90 Labeled Data

Data that includes both input and known output values.

Real-world use:

Used to train supervised learning models, like image datasets with correct classifications or customer data with churn outcomes.

91 Latent Variable

A variable that is not directly observed but inferred from other variables.

Real-world use:

Used in topic modeling of documents, where hidden topics are inferred from word patterns, or in customer segmentation based on purchasing behavior.

92 Learning Rate

The step size used during optimization.

Real-world use:

Used to control speed of training in neural networks, balancing between fast convergence and stable learning (typically 0.001-0.1).

Common mistake:

Setting learning rate too high (causing instability) or too low (causing extremely slow training).

96 Loss Function

A function that measures model error.

Real-world use:

Used to train models by minimizing error, like mean squared error for regression or cross-entropy loss for classification tasks.

121 Parameter

A variable that the model learns during training.

Real-world use:

Used in adjusting weights in neural networks, determining how strongly each input feature influences the final prediction in image recognition or text classification.

122 PCA (Principal Component Analysis)

A dimensionality reduction technique that transforms features into components.

Real-world use:

Used to visualize high-dimensional data, reducing thousands of gene expression features to 2-3 components for cancer research visualization.

123 Perceptron

The simplest type of neural network unit.

Real-world use:

Used in early pattern recognition, serving as the foundation for modern neural networks and linear classification tasks.

125 Pipeline

A sequence of data processing components.

Real-world use:

Used to streamline ML workflows, automatically processing data from cleaning through feature engineering to model training and prediction.

129 Preprocessing

Data preparation steps before training a model.

Real-world use:

Used to remove noise from sensor data, normalize features, handle missing values, and encode categorical variables before model training.

128 Predictive Modeling

Using data to build a model that can predict outcomes.

Real-world use:

Used in credit scoring and churn prediction, helping banks assess loan risk and companies identify customers likely to cancel subscriptions.

137 Sampling

Selecting a subset of data points from a larger set.

Real-world use:

Used in reducing dataset size for training when working with massive datasets, or creating representative samples for statistical analysis.

124 Performance Metric

A quantitative measure used to assess model performance.

Real-world use:

Used in comparing models (e.g., F1, accuracy), enabling data scientists to select the best performing model for production deployment.

126 Poisson Distribution

A distribution that models rare events in a fixed interval.

Real-world use:

Used in predicting server failures or call arrivals, modeling events that occur independently at a constant average rate.

127 Precision

The ratio of true positives to predicted positives.

Real-world use:

Used when false positives are costly, like in spam filtering where marking legitimate emails as spam frustrates users.

Common mistake:

Optimizing for precision alone without considering recall, potentially missing important positive cases.

130 Probability

A measure of the likelihood of an event.

Real-world use:

Used in probabilistic forecasting, risk assessment, and uncertainty quantification in machine learning predictions.

131 Recall

The ratio of true positives to all actual positives.

Real-world use:

Used when missing a positive is costly (e.g., disease detection), ensuring most actual cases are identified even if some false positives occur.

Common mistake:

Confusing recall with precision or not understanding the precision-recall tradeoff in model optimization.

133 Regression

A predictive modeling technique for continuous outcomes.

Real-world use:

Used to forecast revenue, predict house prices, estimate sales figures, and other continuous numerical predictions.

135 Residual

The difference between predicted and actual values.

Real-world use:

Used in diagnosing regression models, analyzing residual patterns to identify model assumptions violations or areas for improvement.

136 ROC Curve

A graph showing the true positive rate vs. false positive rate.

Real-world use:

Used to evaluate binary classifiers across different threshold settings, helping optimize the tradeoff between sensitivity and specificity.

139 Sensitivity

Another term for recall.

Real-world use:

Used in medical diagnosis testing to measure how well a test identifies patients who actually have the disease.

140 SGD (Stochastic Gradient Descent)

An optimization algorithm that updates weights for each data sample.

Real-world use:

Used to train deep learning models efficiently, updating parameters more frequently than batch gradient descent for faster convergence.

141 Standard Deviation

A measure of data spread around the mean.

Real-world use:

Used in risk modeling to quantify volatility in financial markets, portfolio management, and quality control processes.

143 Support Vector Machine (SVM)

A classifier that finds the optimal separating hyperplane.

Real-world use:

Used in text classification, image recognition, and bioinformatics where clear decision boundaries between classes are important.

Core Algorithms

132 Recurrent Neural Network (RNN)

A neural network that processes sequences by looping over data.

Real-world use:

Used in time series forecasting and language modeling, processing sequential data like stock prices, weather patterns, or natural language text.

Common mistake:

Using basic RNNs for very long sequences where vanishing gradient problems make LSTM or GRU more appropriate.

Learning Paradigms

134 Reinforcement Learning

A learning paradigm where agents learn by reward and punishment.

Real-world use:

Used in robotics and game AI, enabling systems to learn optimal strategies through trial and error, like AlphaGo or autonomous vehicle navigation.

138 Semi-Supervised Learning

Training a model on a small labeled set and a large unlabeled set.

Real-world use:

Used when labeling data is expensive, like medical image analysis where expert annotations are costly but raw images are abundant.

142 Supervised Learning

Training a model using labeled data.

Real-world use:

Used in image classification, spam detection, and medical diagnosis where models learn from examples with known correct answers.

145 Underfitting

When a model is too simple to capture patterns in the data.

Real-world use:

Seen when both training and test error are high, like using linear regression for clearly non-linear relationships in stock price prediction.

Common mistake:

Assuming more complexity always improves performance without considering the underlying data patterns.

NLP & Text Processing

144 Tokenization

Splitting text into individual words or symbols.

Real-world use:

Used in NLP preprocessing to break down sentences into analyzable units for sentiment analysis, machine translation, and chatbot development.

Metrics & Measures

81 Jaccard Similarity

A statistic used for comparing the similarity of sample sets.

Real-world use:

Used in text analysis and clustering to measure similarity between documents based on shared words or in recommendation systems for user similarity.

82 Joint Probability

The probability of two events occurring together.

Real-world use:

Used in Naive Bayes classifiers to calculate the probability of multiple features occurring together in spam detection or medical diagnosis.

88 Kurtosis

A measure of the "tailedness" of the probability distribution.

Real-world use:

Used in analyzing stock return distributions to understand risk, where high kurtosis indicates more extreme price movements.

98 Manhattan Distance

A distance metric based on grid-like movement.

Real-world use:

Used in KNN when measuring city-block differences, like calculating taxi distances in urban route planning or clustering categorical data.

99 Marginal Probability

The probability of a single event occurring.

Real-world use:

Used in Bayes' theorem calculations for spam filtering, medical diagnosis, and other probabilistic models to understand individual event likelihood.

100 Mean Absolute Error (MAE)

The average of absolute differences between predictions and actual values.

Real-world use:

Used to evaluate regression models in sales forecasting, providing an interpretable measure of average prediction error in original units.

Optimization & Training

63 Gradient Descent

An optimization algorithm to minimize loss by updating model weights.

Real-world use:

Used in training all neural networks, iteratively adjusting weights to minimize prediction errors in tasks from image recognition to language translation.

65 Grid Search

A method to find the best combination of hyperparameters.

Real-world use:

Used to optimize model settings for maximum accuracy by systematically testing combinations of learning rates, regularization values, and network architectures.

70 Hyperparameter

A configuration value set before training a model.

Real-world use:

Used to tune learning rate, batch size, number of layers, and regularization strength to optimize model performance for specific tasks.

71 Hyperparameter Tuning

The process of choosing the best hyperparameters for a model.

Real-world use:

Used in improving model performance by systematically testing different configurations to find optimal settings for specific datasets and tasks.

80 Iteration

A single update step during model training.

Real-world use:

Used in each step of gradient descent, where the model processes a batch of data and updates weights based on the calculated error.

Specialized Applications

64 Graph Neural Network (GNN)

A neural network designed to operate on graph structures.

Real-world use:

Used in social network analysis and recommendation engines, analyzing relationships between users, products, or molecular structures.

72 Image Classification

The task of assigning a label to an image.

Real-world use:

Used in identifying diseases in medical imaging, quality control in manufacturing, and content moderation on social media platforms.

79 Intersection over Union (IoU)

A metric to evaluate object detection accuracy.

Real-world use:

Used in comparing predicted and ground-truth bounding boxes in autonomous vehicle systems to measure how accurately objects are detected.

Techniques & Methods

49 Exploratory Data Analysis (EDA)

The process of summarizing the main characteristics of data.

Real-world use:

Used to detect patterns and anomalies visually through histograms, scatter plots, and correlation matrices before building models.

50 Extrapolation

Predicting beyond the range of observed data.

Real-world use:

Used in forecasting future sales based on historical trends, predicting population growth, or estimating stock prices beyond training data range.

Common mistake:

Extrapolating too far beyond training data range, leading to unreliable predictions due to unseen patterns.

51 Feature

A measurable input property of the data.

Real-world use:

Used as an input in predictive models, like age, income, and location in house price prediction or pixel values in image classification.

52 Feature Engineering

Creating or modifying features to improve model performance.

Real-world use:

Used to extract date parts from timestamps (day, month, year) for sales forecasting or creating interaction features for better predictions.

53 Feature Importance

A score indicating how much a feature influences prediction.

Real-world use:

Used to interpret tree-based models, identifying which factors most influence loan approvals or medical diagnoses for stakeholder understanding.

54 Feature Selection

Choosing the most relevant features for a model.

Real-world use:

Used to reduce overfitting and improve model speed by selecting only the most predictive variables from hundreds of potential features.

57 Fine-Tuning

Adjusting a pretrained model for a new task.

Real-world use:

Used to adapt BERT for sentiment analysis or fine-tune image classification models for specific domains like medical imaging.

59 Frequency Encoding

A method of encoding categorical data using frequency counts.

Real-world use:

Used in converting nominal values for ML models, replacing city names with their occurrence frequency in the dataset for better model performance.

60 Function Approximation

Estimating an unknown function that best fits input/output pairs.

Real-world use:

Used in regression modeling to approximate the relationship between house features and prices, or between advertising spend and sales revenue.

Neural Networks & Deep Learning

3 Activation Function

A function in neural networks that introduces non-linearity.

Real-world use:

Used in deep learning models like ReLU in image recognition, enabling networks to learn complex patterns like edges and shapes.

6 ANN (Artificial Neural Network)

Model inspired by biological brains for pattern recognition.

Real-world use:

Used in handwriting recognition for check processing at banks and postal services for address reading.

12 Backpropagation

An algorithm for training neural networks by updating weights based on errors.

Real-world use:

Used in deep learning for optimization, enabling networks to learn from mistakes and improve performance on tasks like image classification.

14 Batch Size

The number of training examples used in one iteration.

Real-world use:

Used in configuring neural network training, balancing memory usage and training stability (e.g., batch size of 32 for image classification).

Common mistake:

Using batch sizes that are too large for available memory or too small, causing unstable training.

26 Convolutional Layer

A neural network layer for spatial feature detection.

Real-world use:

Used in image classification models to detect edges, textures, and patterns in medical imaging, autonomous vehicles, and photo recognition.

38 Deep Learning

A subset of machine learning with multilayer neural networks.

Real-world use:

Used in voice recognition systems like Siri and Alexa, enabling natural language understanding and speech-to-text conversion.

41 Discriminator

The neural network in a GAN that identifies fake data.

Real-world use:

Used in generative adversarial image tasks, distinguishing between real photos and AI-generated images in deepfake detection systems.

42 Dropout

A regularization technique to prevent overfitting in neural networks.

Real-world use:

Used during training to randomly deactivate neurons, improving generalization in deep learning models for image and text classification.

47 Epoch

One full pass over the entire training dataset.

Real-world use:

Used in training deep neural networks, with models typically requiring 10-100 epochs to converge on tasks like image classification.

55 Feedforward Network

A type of neural network with connections moving in one direction.

Real-world use:

Used in basic image recognition models and tabular data classification, forming the foundation for more complex architectures.

58 Forward Propagation

The process of computing output predictions in a neural network.

Real-world use:

Used in every training iteration and prediction, passing input data through network layers to generate final outputs.

43 Elbow Method

A technique to choose the number of clusters in k-means.

Real-world use:

Used to identify optimal k-value in customer segmentation, finding the "elbow" point where adding more clusters doesn't significantly improve performance.

44 Embedding

A dense vector representation of data.

Real-world use:

Used in NLP to represent words numerically, enabling models to understand semantic relationships like "king - man + woman = queen."

45 Ensemble Learning

Combining multiple models to improve predictions.

Real-world use:

Used in Random Forests and boosting algorithms, combining hundreds of decision trees to create more accurate and robust predictions.

46 Entropy

A measure of disorder or randomness.

Real-world use:

Used to split nodes in decision trees, selecting features that best separate classes by reducing uncertainty in the data.

AI Practitioner Glossary

Fundamentals & Core Concepts

4 AI (Artificial Intelligence)

Machines performing tasks that typically require human intelligence.

Real-world use:

Used in personal assistants like Siri, Alexa, and Google Assistant for voice recognition and natural language processing.

5 Algorithm

A set of rules or calculations used by machines to solve problems.

Real-world use:

Used in route optimization for delivery services like UPS and FedEx to minimize travel time and fuel costs.

10 Attribute

A variable or column in a dataset representing a specific property.

Real-world use:

Used in machine learning features like age, income, and credit score for loan approval models.

7 Anomaly

A data point that deviates significantly from the norm.

Real-world use:

Used in fraud detection to identify unusual spending patterns on credit cards.

16 Bias

Systematic error from incorrect model assumptions.

Real-world use:

Seen in AI models trained on unbalanced datasets, like facial recognition systems performing poorly on underrepresented demographics.

Common mistake:

Assuming all prediction errors are due to bias when some may be due to variance or irreducible error.

20 Categorical Data

Data sorted by category rather than quantity.

Real-world use:

Used to label items like color, gender, or brand in customer segmentation and market research.

Model Evaluation & Performance

21 Precision

The ratio of true positive predictions to all positive predictions made by the model.

Real-world use:

Used in medical diagnosis where false positives are costly - measuring how many cancer diagnoses are actually correct.

Common mistake:

Optimizing for precision alone without considering recall, leading to overly conservative models.

22 Recall

The ratio of true positive predictions to all actual positive cases.

Real-world use:

Critical in fraud detection where missing actual fraud cases is expensive - measuring how many fraudulent transactions are caught.

Common mistake:

Confusing recall with precision or not understanding the precision-recall tradeoff.

23 F1-Score

The harmonic mean of precision and recall, providing a single balanced metric.

Real-world use:

Used when you need to balance precision and recall, like in document classification where both false positives and false negatives matter.

24 Confusion Matrix

A table showing correct vs. predicted classifications for each class.

Real-world use:

Used to analyze model performance in image recognition, showing which objects are commonly misclassified as others.

25 Cross-Validation

A technique for assessing model performance by training and testing on different data subsets.

Real-world use:

Used in machine learning competitions to ensure models generalize well beyond the training data, typically using 5-fold or 10-fold CV.

26 Overfitting

When a model learns training data too specifically and fails to generalize to new data.

Real-world use:

Common in complex models like deep neural networks, where the model memorizes training examples rather than learning patterns.

Common mistake:

Using training accuracy to judge model quality instead of validation performance.

Data & Processing

27 Data Preprocessing

The process of cleaning and transforming raw data before analysis.

Real-world use:

Used in customer analytics to clean purchase data, handle missing values, and standardize formats before building recommendation systems.

28 Feature Engineering

Creating new features or transforming existing ones to improve model performance.

Real-world use:

Used in credit scoring to create features like debt-to-income ratio from separate debt and income columns.

29 Normalization

Scaling data to a standard range, typically 0 to 1.

Real-world use:

Used in neural networks to ensure features like age (0-100) and salary ($20k-$200k) have similar scales for fair comparison.

30 Outlier Detection

Identifying data points that significantly differ from the majority of data.

Real-world use:

Used in network security to identify unusual traffic patterns that might indicate cyber attacks or system breaches.

31 Missing Data

Absent values in a dataset that need to be handled before analysis.

Real-world use:

Common in survey data where respondents skip questions, requiring strategies like mean imputation or deletion before analysis.

Common mistake:

Simply deleting all rows with missing values without considering the impact on data representativeness.

32 Dimensionality Reduction

Reducing the number of features while preserving important information.

Real-world use:

Used in image compression and visualization, reducing thousands of pixel features to key components while maintaining image quality.

29 Data Augmentation

Creating new data by modifying existing data.

Real-world use:

Used in training with limited image datasets, creating variations through rotation, scaling, and cropping to improve model generalization.

30 Data Leakage

When test data information is used during training.

Real-world use:

Leads to overestimated model performance in production, like accidentally including future information in stock price prediction models.

Common mistake:

Applying preprocessing steps before splitting data, causing information from test set to influence training.

31 Data Mining

Discovering patterns in large data sets.

Real-world use:

Used in customer purchase behavior analysis to identify shopping patterns, seasonal trends, and product associations for inventory management.

32 Data Normalization

Scaling values to a standard range.

Real-world use:

Used to prepare features for neural networks, ensuring variables like age (0-100) and income ($0-$200k) contribute equally to learning.

33 Data Pipeline

A series of data processing steps.

Real-world use:

Used to automate ETL tasks in data engineering, moving data from databases through cleaning, transformation, and loading into analytics systems.

35 Data Wrangling

Restructuring and cleaning data for use.

Real-world use:

Used to prepare messy survey data for modeling, handling inconsistent formats, missing values, and combining multiple data sources.

Advanced Concepts

33 Clustering

Grouping similar data points together without predefined labels.

Real-world use:

Used in customer segmentation to group customers by purchasing behavior, enabling targeted marketing campaigns.

34 Reinforcement Learning

Learning through trial and error using rewards and penalties.

Real-world use:

Used in game AI like AlphaGo, autonomous vehicles, and recommendation systems that learn from user interactions.

35 Transfer Learning

Using knowledge from one task to improve performance on a related task.

Real-world use:

Used in medical imaging where models trained on general images are fine-tuned for X-ray analysis, reducing training time and data needs.

36 Ensemble Learning

Combining multiple models to create a stronger predictor.

Real-world use:

Used in Kaggle competitions and production systems, combining decision trees, neural networks, and linear models for better predictions.

37 Hyperparameter Tuning

Optimizing model settings that aren't learned during training.

Real-world use:

Used to find optimal learning rates, network architectures, and regularization parameters for better model performance.

38 NLP (Natural Language Processing)

AI field focused on understanding and generating human language.

Real-world use:

Used in chatbots, language translation, sentiment analysis, and document summarization across industries.

Statistics & Testing

1 A/B Test

A hypothesis test comparing two variants of a single variable to determine which performs better.

Real-world use:

Used in email marketing to test subject lines, comparing open rates between "Free Shipping" vs "50% Off" headlines.

8 ANOVA

Analysis of variance, used to compare means of three or more groups.

Real-world use:

Used in testing customer engagement across multiple app versions to determine which interface design performs best.

9 Arithmetic Mean

The average of a list of numbers.

Real-world use:

Used to calculate average student scores, sales performance, or website loading times for performance reporting.

15 Bayesian Inference

A method for updating probability estimates using new data.

Real-world use:

Used in filtering spam with changing signals, adapting to new spam techniques by updating probability models.

22 Chi-Squared Test

A test comparing observed results to expected ones.

Real-world use:

Used to test independence of variables in surveys, like determining if gender and product preference are related.

24 Coefficient of Determination (R²)

Shows how well a regression line fits data.

Real-world use:

Used to evaluate predictive accuracy in sales forecasting, where R² = 0.85 means 85% of sales variation is explained by the model.

Common mistake:

Assuming high R² always means a good model, when it could indicate overfitting or spurious correlations.

Machine Learning

2 Accuracy

The ratio of correct predictions to total predictions.

Real-world use:

Used in spam filters to measure correct classification, tracking how often emails are properly labeled as spam or legitimate.

Common mistake:

Confusing high accuracy with good model performance in imbalanced datasets where most examples belong to one class.

11 AUC (Area Under Curve)

Measures the performance of a binary classification model.

Real-world use:

Used to evaluate model discrimination ability in medical diagnosis, measuring how well a model distinguishes between disease and non-disease cases.

13 Baseline

A simple reference model used for performance comparison.

Real-world use:

Used in evaluating model improvements, like comparing a complex recommendation system against simply recommending the most popular items.

17 Binary Classification

Categorizing inputs into two groups.

Real-world use:

Used in sentiment analysis to classify customer reviews as positive vs. negative, or in medical screening for disease vs. no disease.

18 Boosting

Combining weak models to create a stronger model.

Real-world use:

Used in XGBoost or AdaBoost for fraud detection, combining multiple simple decision rules to create highly accurate predictions.

19 CART

Classification and regression trees used for prediction.

Real-world use:

Used in decision tree models for loan approval, creating interpretable rules like "if income > $50k and credit score > 700, then approve."

21 Centroid

The central point of a cluster in data.

Real-world use:

Used in k-means clustering algorithms to represent the center of customer segments or product categories.

23 Clustering

Grouping similar data without predefined labels.

Real-world use:

Used in customer segmentation to identify distinct groups based on purchasing behavior, demographics, and preferences for targeted marketing.

27 Cost Function

A measure of model prediction error.

Real-world use:

Used in gradient descent optimization to minimize prediction errors, like mean squared error in housing price prediction models.

36 Decision Boundary

A surface separating different predicted classes.

Real-world use:

Used in visualizing classification models, showing how a model separates spam from legitimate emails in feature space.

37 Decision Tree

A model that splits data using rules based on features.

Real-world use:

Used in loan approval systems, creating interpretable decision paths like "if credit score > 700 and income > $50K, then approve loan."

40 Discretization

Converting continuous variables into categorical ones.

Real-world use:

Used to group ages into ranges like "18-25, 26-35, 36-50" for market research and demographic analysis.

Neural Networks & Deep Learning

3 Activation Function

A function in neural networks that introduces non-linearity.

Real-world use:

Used in deep learning models like ReLU in image recognition, enabling networks to learn complex patterns like edges and shapes.

6 ANN (Artificial Neural Network)

Model inspired by biological brains for pattern recognition.

Real-world use:

Used in handwriting recognition for check processing at banks and postal services for address reading.

12 Backpropagation

An algorithm for training neural networks by updating weights based on errors.

Real-world use:

Used in deep learning for optimization, enabling networks to learn from mistakes and improve performance on tasks like image classification.

14 Batch Size

The number of training examples used in one iteration.

Real-world use:

Used in configuring neural network training, balancing memory usage and training stability (e.g., batch size of 32 for image classification).

Common mistake:

Using batch sizes that are too large for available memory or too small, causing unstable training.

Matthew Sutherland 5/22/25 Matthew Sutherland 5/22/25

Prompt for Graphic Artist: Anime Cyber Samurai

The Evolution of the Digital Warrior

The concept of tech-enhanced warriors fighting in dystopian futures isn't new—classics like "Ghost in the Shell" and "Akira" pioneered this territory decades ago. However, modern interpretations have evolved significantly, integrating deeper philosophical questions about humanity's relationship with artificial intelligence.

Today's cyber samurai characters are not just battling external AI threats; they're often exploring the boundaries between human consciousness and digital existence. Their weapons—traditional katanas infused with energy fields or digital enhancements—symbolize this duality, bridging ancient craft with bleeding-edge innovation.

This single prompt was used verbatim on multiple platforms with very different results. Enjoy!

Prompt for Graphic Artist:
"Create a dynamic anime-style illustration of a Japanese girl in the midst of a high-stakes battle against a futuristic AI entity.
* Character: Fierce expression, wearing a sleek, tech-enhanced combat uniform blending traditional Japanese elements (e.g., katana, obi sash) with cybernetic accents (glowing circuits, holographic visor). Her hair flows dramatically, electrified with neon streaks.
* Setting: A dystopian cyber-cityscape with crumbling neon skyscrapers, floating holograms, and rogue AI drones swarming like mechanical insects.
* Action: She leaps mid-air, katana charged with energy, slicing through a towering AI core manifesting as a fractal-like digital monster with glowing red eyes and tendrils of corrupted data.
* Color Palette: Vivid neons (cyan, magenta, electric purple) contrasted against dark metallic grays and smoky shadows.
* Style: Bold linework, cinematic lighting, and hyper-detailed mechanical designs reminiscent of Ghost in the Shell or Akira."
Update note: A few words had to be changed in order to meet the terms and conditions of the individual sources. ie. Girl changed to female.