Wasserstein Distance
A metric used in optimal transport theory to measure the distance between probability distributions.
Types of Wasserstein Distance
- Wasserstein-1 Distance (Earth Mover’s Distance) - Measures minimum effort to transform one distribution into another.
- Wasserstein-2 Distance - Considers the squared Euclidean distance for optimal transport.
Example
Used in Generative Adversarial Networks (WGANs) for stable training.
Weakly Supervised Learning
A learning paradigm where models are trained on partially labeled or noisy data.
Types of Weak Supervision
- Incomplete Supervision - Uses limited labeled data with a large unlabeled set.
- Inexact Supervision - Labels provide only partial information (e.g., bounding boxes instead of segmentation masks).
Example
Used in medical imaging where only rough annotations are available.
Weight Decay
A regularization technique that reduces overfitting by penalizing large weights in neural networks.
Types of Weight Decay
- L1 Regularization - Shrinks weights to zero, leading to sparse models.
- L2 Regularization - Penalizes large weights without forcing them to zero.
Example
Used in deep learning models like CNNs and RNNs to improve generalization.
Weighted Average Precision (WAP)
A metric that computes the precision of a model by considering class imbalances in multi-class classification tasks.
Types of Weighted Precision
- Micro-Averaged Precision - Computes precision globally across all instances.
- Macro-Averaged Precision - Computes precision per class and averages them.
Example
Used in evaluating imbalanced datasets, such as fraud detection.
Weight Initialization
A process of setting initial values for neural network weights to stabilize learning.
Types of Weight Initialization
- Xavier Initialization - Scales weights for balanced gradients.
- He Initialization - Optimized for ReLU-based networks.
Example
Used in deep learning architectures like ResNets and Transformers.
Wasserstein GAN (WGAN)
A variation of GANs that stabilizes training using Wasserstein distance instead of JS divergence.
Types of WGAN
- WGAN - Uses weight clipping for stability.
- WGAN-GP - Uses gradient penalty instead of weight clipping.
Example
Used in realistic image synthesis and style transfer.
Types of Wavelet Transform
- Continuous Wavelet Transform (CWT) - Provides detailed time-frequency analysis.
- Discrete Wavelet Transform (DWT) - Used in data compression and denoising.
Example
Used in feature extraction for time-series and image processing.
Web Scraping for Machine Learning
A method of collecting data from the web to train machine learning models.
Types of Web Scraping
- Static Scraping - Extracts data from fixed web pages.
- Dynamic Scraping - Uses automation to interact with web elements.
Example
Used in sentiment analysis and price prediction models.
Weighted Ensemble Learning
A technique that assigns different importance levels to models in an ensemble for improved performance.
Types of Weighted Ensemble Learning
- Bagging - Assigns equal weight to models (e.g., Random Forest).
- Boosting - Adjusts weights based on model errors (e.g., AdaBoost, XGBoost).
Example
Used in Kaggle competitions for robust model performance.
Windowing in Time Series
A method of segmenting time-series data into overlapping or non-overlapping windows for machine learning.
Types of Windowing
- Fixed-Size Windowing - Uses a constant-sized window.
- Dynamic Windowing - Adjusts window size based on data patterns.
Example
Used in forecasting models like LSTMs and ARIMA.
Word Embeddings
A representation of words as numerical vectors in a continuous space to capture semantic meaning.
Types of Word Embeddings
- Word2Vec - Predicts word relationships using skip-gram or CBOW models.
- GloVe - Captures word co-occurrence statistics in a vector space.
Example
Used in NLP tasks like sentiment analysis and chatbots.
Weak Learner
A model that performs slightly better than random guessing, often used in ensemble learning.
Types of Weak Learners
- Decision Stumps - Single-level decision trees.
- Naïve Bayes - A probabilistic weak classifier.
Example
Used in boosting algorithms like AdaBoost and XGBoost.
Whitened Data
Data that has been transformed to have zero mean and unit variance to improve learning stability.
Types of Whitening
- PCA Whitening - Uses principal component analysis to decorrelate features.
- ZCA Whitening - Preserves spatial structure while normalizing.
Example
Used in deep learning for input normalization.
Warm Start
A training technique where previous model weights are used to accelerate learning instead of starting from scratch.
Types of Warm Start
- Fine-Tuning - Adjusting a pre-trained model to a new task.
- Continual Learning - Retaining learned knowledge across training iterations.
Example
Used in transfer learning with models like BERT and ResNet.
Weight Sharing
A technique where the same set of weights is used across multiple layers or inputs to reduce model complexity.
Types of Weight Sharing
- Convolutional Neural Networks (CNNs) - Uses shared filters to detect patterns.
- Recurrent Neural Networks (RNNs) - Shares weights across time steps.
Example
Used in deep learning architectures for efficient parameter usage.
Wasserstein Loss
A loss function used in WGANs to measure distribution distance and stabilize training.
Types of Wasserstein Loss
- Wasserstein-1 Loss - Measures transportation cost.
- Gradient Penalty Loss - Enhances stability in WGAN-GP.
Example
Used in training Generative Adversarial Networks (GANs).
Weak Annotation
Labels that are noisy, incomplete, or approximate rather than fully supervised.
Types of Weak Annotation
- Soft Labels - Probabilistic labels instead of binary labels.
- Bounding Boxes - Instead of pixel-perfect segmentation.
Example
Used in self-supervised learning to leverage large datasets.
Wide and Deep Learning
A model combining linear models (wide) with deep neural networks (deep) for better performance.
Types of Wide and Deep Models
- Wide Component - Memorization of past patterns.
- Deep Component - Generalization to new data.
Example
Used in recommendation systems like Google Play Store.
Window Function
A function that selects a subset of data points within a specified window for analysis.
Types of Window Functions
- Hamming Window - Used for smooth signal processing.
- Hanning Window - Reduces spectral leakage in FFT.
Example
Used in time-series forecasting and signal processing.
Weighted Feature Selection
A method that assigns different importance levels to features before model training.
Types of Weighted Feature Selection
- Filter-Based Selection - Uses statistical scores like Chi-Square.
- Embedded Methods - Uses models like Lasso for selection.
Example
Used in high-dimensional datasets like bioinformatics.
Weighted Sampling
A technique where different data points are sampled based on assigned probabilities to handle class imbalance.
Types of Weighted Sampling
- Oversampling - Increases the presence of minority classes.
- Undersampling - Reduces the majority class to balance the dataset.
Example
Used in fraud detection to balance imbalanced datasets.
Weight Decay
A regularization technique that penalizes large weight values to prevent overfitting.
Types of Weight Decay
- L1 Regularization - Promotes sparsity by adding an absolute penalty.
- L2 Regularization - Adds a squared penalty to encourage small weights.
Example
Applied in deep learning models like neural networks to improve generalization.
Weight Normalization
A reparameterization technique that normalizes weight vectors to accelerate training.
Types of Weight Normalization
- Scale-Based Normalization - Separates magnitude and direction of weights.
- Batch Normalization Alternative - Normalizes weights instead of activations.
Example
Used in deep learning models for faster convergence.
Wasserstein Distance
A metric used to measure the distance between probability distributions, often applied in GANs.
Types of Wasserstein Distance
- W1 Distance - Measures earth mover’s distance.
- Sinkhorn Distance - Adds entropy regularization for smoother computation.
Example
Used in WGANs to stabilize training.
Weak Supervision
A training method that leverages noisy, limited, or imprecise labels to improve model learning.
Types of Weak Supervision
- Distant Supervision - Uses external knowledge bases to generate labels.
- Multiple Instance Learning - Trains models with uncertain labeling.
Example
Used in NLP and medical image classification where labeled data is scarce.
WGAN (Wasserstein GAN)
A variant of GANs that uses Wasserstein distance instead of traditional loss functions to improve training stability.
Types of WGAN
- WGAN-GP - Adds gradient penalty for better convergence.
- Original WGAN - Uses weight clipping to enforce Lipschitz continuity.
Example
Used in image generation and synthetic data creation.
Workload Distribution
A method for balancing computational tasks across multiple processors or systems.
Types of Workload Distribution
- Static Distribution - Predefined allocation of tasks.
- Dynamic Distribution - Adaptive task allocation based on real-time conditions.
Example
Used in cloud computing and parallel machine learning training.
Window Size in NLP
The number of words or tokens considered at a time for context-based processing.
Types of Window Sizes
- Fixed Window - A predefined number of words.
- Dynamic Window - Adjusts based on sentence structure.
Example
Used in Word2Vec and n-gram models for text analysis.
Weighted Loss Function
A loss function that assigns different importance levels to different samples or classes.
Types of Weighted Loss Functions
- Class-Weighted Loss - Adjusts loss for imbalanced datasets.
- Sample-Weighted Loss - Assigns importance based on sample difficulty.
Example
Used in deep learning models dealing with class imbalances.
Weakly Labeled Data
Data where labels are incomplete, ambiguous, or approximate rather than fully annotated.
Types of Weakly Labeled Data
- Partially Labeled Data - Some samples have labels, others do not.
- Noisy Labels - Contains incorrect or uncertain labeling.
Example
Used in semi-supervised learning for training AI models with limited supervision.
White Box Model
A machine learning model whose internal workings are transparent and explainable.
Types of White Box Models
- Linear Models - Easy to interpret, such as linear regression.
- Decision Trees - Provide clear decision paths.
Example
Used in healthcare AI for explainable medical diagnoses.
Types of Wavelet Transform
- Continuous Wavelet Transform (CWT) - Captures fine-grained signal variations.
- Discrete Wavelet Transform (DWT) - Used for dimensionality reduction.
Example
Applied in speech recognition and financial forecasting.
Weak Learner
A model that performs slightly better than random guessing but can be boosted for improved performance.
Types of Weak Learners
- Decision Stumps - Single-level decision trees.
- Naïve Bayes - Assumes feature independence.
Example
Used in boosting algorithms like AdaBoost and Gradient Boosting.
Wasserstein Autoencoder (WAE)
A type of autoencoder that minimizes Wasserstein distance to improve generative modeling.
Types of WAE
- WAE-GAN - Uses adversarial training.
- WAE-MMD - Uses Maximum Mean Discrepancy for distribution alignment.
Example
Used for generating high-quality synthetic data.
Weight Initialization
The process of setting initial weight values in neural networks to ensure stable training.
Types of Weight Initialization
- Xavier Initialization - Scales weights based on the number of neurons.
- He Initialization - Optimized for ReLU activations.
Example
Used in deep learning architectures to prevent vanishing gradients.
Weighted Feature Selection
A method of selecting features based on their importance scores to improve model efficiency.
Types of Weighted Feature Selection
- Information Gain - Measures feature relevance using entropy.
- Mutual Information - Evaluates dependencies between features.
Example
Applied in text classification to eliminate irrelevant words.
Windowing Technique in ML
A method used in data processing where a subset of data is analyzed within a moving window.
Types of Windowing
- Fixed Window - Uses a constant window size.
- Adaptive Window - Adjusts window size based on data changes.
Example
Used in time-series forecasting and speech recognition.
Word Embedding
A technique to represent words as continuous vectors in a high-dimensional space.
Types of Word Embeddings
- Word2Vec - Predicts words based on surrounding context.
- GloVe - Uses co-occurrence statistics for embedding.
Example
Used in NLP models for semantic understanding.
Wrapper Method in Feature Selection
A feature selection technique that evaluates subsets of features using model performance.
Types of Wrapper Methods
- Forward Selection - Adds features iteratively.
- Backward Elimination - Removes features iteratively.
Example
Used in predictive modeling to improve accuracy.
Weighted Kernel in SVM
A method in Support Vector Machines (SVM) where different weights are assigned to kernel functions for optimization.
Types of Weighted Kernels
- Polynomial Kernel - Used for complex relationships.
- Gaussian Kernel - Used for non-linear decision boundaries.
Example
Applied in SVM-based image classification.
Machine Learning (ML)
ML is a subset of AI that enables machines to learn patterns from data and make predictions or decisions without explicit programming.
Types of ML
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning
Example
Spam detection in emails using classification models.
Deep Learning (DL)
DL is a subset of ML that uses artificial neural networks to process complex data and perform high-level computations.
Example
Image recognition in self-driving cars.
Generative AI (Gen AI)
Gen AI refers to AI models that generate new content, including text, images, and code, using trained knowledge bases.
Example
AI models like ChatGPT and Stable Diffusion that generate text and images.