Menü aufrufen
Toggle preferences menu
Persönliches Menü aufrufen
Nicht angemeldet
Ihre IP-Adresse wird öffentlich sichtbar sein, wenn Sie Änderungen vornehmen.

Machine Learning Glossary/O

Aus VELEVO®.WIKI

Verfasst von / Written by Sebastian F. Genter


O

objective

In machine learning, an objective refers to the specific goal or target metric that an algorithm is designed to optimize during training. Objectives guide the learning process by quantifying how well the model performs on a given task. Common objectives include minimizing prediction errors (for regression) or maximizing classification accuracy.

objective function

The mathematical formulation that defines what the model aims to optimize. Also known as a loss function or cost function, it quantifies the difference between predicted and actual values. Different machine learning tasks require different objective functions:

  • Regression often uses mean squared error
  • Classification might use cross-entropy loss
  • Reinforcement learning may use reward maximization

oblique condition

In decision tree algorithms, a condition that involves multiple features simultaneously. Unlike axis-aligned conditions that test a single feature (e.g., "age > 30"), oblique conditions combine features (e.g., "2*height + weight > 300"). These can create more complex decision boundaries but may increase computational complexity.

offline

Refers to processes that occur without real-time requirements, typically involving pre-computation or batch processing. In machine learning contexts:

  • Offline training occurs on static datasets
  • Offline evaluation assesses model performance
  • Offline inference generates predictions in advance

offline inference

A prediction approach where models generate outputs in batches before they're needed, storing results for later use. Benefits include:

  • Reduced computational load during serving
  • Consistent response times
  • Ability to handle spikes in demand

Common in recommendation systems and weather forecasting where predictions can be pre-computed.

one-hot encoding

A technique for representing categorical variables as binary vectors where:

  • Each category becomes a binary feature
  • Only one feature is "hot" (1) per sample
  • All others are "cold" (0)

Example for animal types:

  • Dog: [1, 0, 0]
  • Cat: [0, 1, 0]
  • Bird: [0, 0, 1]

Advantages include simplicity and compatibility with many algorithms, though it can create high-dimensional spaces for variables with many categories.

one-shot learning

A learning paradigm where models make predictions after seeing just one or very few examples of each class. Particularly valuable when:

  • Training data is extremely scarce
  • New categories emerge frequently
  • Rapid adaptation is required

Common approaches include:

  • Metric learning techniques
  • Memory-augmented neural networks
  • Transfer learning from related domains

one-shot prompting

In large language models, providing a single example within the prompt to demonstrate the desired response format. Structure typically includes: 1. Task description 2. Single example input-output pair 3. New input to process

Example:

Translate English to French:

Hello → Bonjour

Goodbye →


one-vs.-all

A multi-class classification strategy that trains N binary classifiers for N classes, where each classifier distinguishes one class from all others. For 3 classes (A,B,C), this would involve: 1. Classifier 1: A vs. (B or C) 2. Classifier 2: B vs. (A or C) 3. Classifier 3: C vs. (A or B)

Final predictions combine results from all classifiers.

online

Refers to real-time or continuous processes in machine learning systems. Characteristics include:

  • Immediate response requirements
  • Continuous data streams
  • Adaptive model updates

Contrasts with offline/batch processing approaches.

online inference

Generating model predictions in real-time as requests arrive. Key aspects:

  • Low latency requirements
  • Direct interaction with users/applications
  • Dynamic input processing

Common in applications like:

  • Fraud detection
  • Chatbots
  • Real-time recommendations

operation (op)

In computational frameworks like TensorFlow, an operation represents a node in the computation graph that:

  • Takes tensors as inputs
  • Performs specific computations
  • Produces tensors as outputs

Examples include:

  • Mathematical operations (add, multiply)
  • Neural network layers (convolution, pooling)
  • Control flow operations (loops, conditionals)

Optax

A gradient processing and optimization library for JAX that provides:

  • Composable gradient transformations
  • Popular optimization algorithms
  • Utilities for machine learning research

Key features:

  • Clean, functional API design
  • Easy combination of optimization components
  • Accelerated computation on GPUs/TPUs

optimizer

Algorithms that adjust model parameters to minimize the objective function. Common optimizers include:

  • Stochastic Gradient Descent (SGD)
  • Adam (Adaptive Moment Estimation)
  • RMSprop (Root Mean Square Propagation)

Key considerations when choosing optimizers:

  • Convergence speed
  • Memory requirements
  • Handling of sparse gradients
  • Robustness to hyperparameters

out-group homogeneity bias

A cognitive bias where individuals perceive members of other groups as more similar to each other than they actually are. In ML contexts, this can lead to:

  • Oversimplified feature representations
  • Reduced model performance on minority groups
  • Unfair treatment of diverse populations

Mitigation strategies include:

  • Diverse dataset collection
  • Careful feature engineering
  • Fairness-aware model evaluation

outlier detection

Techniques for identifying data points that deviate significantly from the majority of the dataset. Approaches include:

  • Statistical methods (Z-scores, IQR)
  • Density-based techniques (DBSCAN)
  • Machine learning models (Isolation Forests)

Applications span:

  • Fraud detection
  • Quality control
  • Anomaly monitoring

outliers

Data points that differ markedly from other observations in the dataset. Can arise from:

  • Measurement errors
  • Rare events
  • Data corruption

Impact on ML models:

  • May distort statistical measures
  • Can disproportionately influence model training
  • Sometimes represent valuable edge cases

out-of-bag evaluation (OOB evaluation)

A validation method for ensemble models (especially random forests) that:

  • Uses samples not selected in bootstrap aggregation
  • Provides unbiased performance estimates
  • Doesn't require separate validation data

Calculation process: 1. For each tree, identify unsampled instances 2. Aggregate predictions for these instances 3. Compare to true labels

output layer

The final layer of a neural network that produces predictions. Characteristics vary by task:

  • Regression: Single node with linear activation
  • Binary classification: Single node with sigmoid
  • Multi-class: Multiple nodes with softmax

Design considerations include:

  • Dimensionality matching output space
  • Appropriate activation functions
  • Connection patterns to previous layers

overfitting

When a model learns patterns specific to the training data that don't generalize to new data. Indicators include:

  • High training accuracy but low validation accuracy
  • Complex decision boundaries
  • Sensitivity to small input changes

Prevention techniques:

  • Regularization (L1/L2)
  • Early stopping
  • Data augmentation
  • Simplifying model architecture

oversampling

A technique for handling imbalanced datasets by increasing representation of minority classes. Methods include:

  • Random duplication
  • SMOTE (Synthetic Minority Oversampling)
  • ADASYN (Adaptive Synthetic Sampling)

Balances with undersampling to:

  • Improve model performance on rare classes
  • Prevent classifier bias
  • Enhance decision boundaries