Verfasst von / Written by Sebastian F. Genter
R
RAG (Retrieval-Augmented Generation)
A hybrid AI architecture that combines:
- Information retrieval from external knowledge sources
- Generative language model capabilities
Key components: 1. Retrieval system finds relevant documents 2. Generator incorporates retrieved information into responses
Benefits:
- Reduces hallucinations
- Enables fact-checking against sources
- Allows knowledge updates without retraining
random forest
An ensemble learning method that operates by:
- Constructing multiple decision trees
- Outputting the mode (classification) or mean prediction (regression) of individual trees
Advantages:
- Handles high-dimensional data well
- Resistant to overfitting
- Provides feature importance metrics
random policy
In reinforcement learning, a strategy that:
- Selects actions uniformly at random
- Serves as baseline for comparison
- Useful for initial exploration
rank (ordinality)
The position of an item in an ordered sequence. Important for:
- Ranking problems (search results, recommendations)
- Ordinal regression tasks
- Evaluation metrics like NDCG
rank (Tensor)
The number of dimensions in a tensor:
- Scalar: rank 0 (e.g., 5)
- Vector: rank 1 (e.g., [1,2,3])
- Matrix: rank 2 (e.g., [[1,2],[3,4]])
- Higher-order tensors: rank 3+
ranking
The process of ordering items by relevance/importance. Applications include:
- Search engine results
- Recommendation systems
- Ad placement
rater
A human evaluator who:
- Labels training data
- Assesses model outputs
- Provides feedback for RLHF
recall
A classification metric measuring:
- True positives / (True positives + False negatives)
- The model's ability to find all relevant instances
recall at k (recall@k)
Evaluation metric measuring:
- Proportion of relevant items in top k results
- Used in recommender systems and IR
recommendation system
Systems that predict user preferences for:
- Products
- Content
- Services
Main approaches:
- Collaborative filtering
- Content-based
- Hybrid
Rectified Linear Unit (ReLU)
A popular activation function:
- f(x) = max(0, x)
- Addresses vanishing gradient problem
- Computationally efficient
recurrent neural network
Neural networks with:
- Temporal connections
- Memory of previous inputs
- Applications in sequence modeling
reference text
Gold standard text used for:
- Model evaluation
- Training supervision
- Quality benchmarking
regression model
Models predicting continuous values:
- Linear regression
- Polynomial regression
- Neural network regression
regularization
Techniques preventing overfitting:
- L1/L2 regularization
- Dropout
- Early stopping
regularization rate
Hyperparameter controlling:
- Strength of regularization
- Tradeoff between fit and simplicity
reinforcement learning (RL)
Learning paradigm where agents:
- Interact with environment
- Receive rewards/penalties
- Optimize long-term return
RLHF (Reinforcement Learning from Human Feedback)
Training process combining:
- Human preference judgments
- Reward modeling
- Policy optimization
replay buffer
In RL, stores:
- Past experiences (state, action, reward)
- Enables experience replay
- Improves sample efficiency
replica
Duplicate components for:
- Parallel processing
- Fault tolerance
- Load balancing
reporting bias
When available data:
- Overrepresents certain phenomena
- Underrepresents others
- Distorts model learning
representation
How data is encoded for:
- Machine processing
- Feature extraction
- Dimensionality reduction
re-ranking
Secondary ranking phase that:
- Refines initial results
- Incorporates additional signals
- Improves final ordering
return
In RL, the cumulative:
- Discounted future rewards
- Objective to maximize
- Measure of policy quality
reward
In RL, the:
- Immediate feedback signal
- Numerical evaluation of actions
- Driver of learning
ridge regularization
L2 regularization that:
- Adds squared magnitude penalty
- Prevents coefficient explosion
- Encourages small weights
RNN
Abbreviation for recurrent neural network
ROC Curve
Graphical plot showing:
- True positive rate vs false positive rate
- Across classification thresholds
- Visualizes tradeoffs
role prompting
Technique where:
- LLM is assigned a specific role
- ("You are a helpful assistant")
- Guides response style
root
In decision trees:
- The initial node
- Contains all training data
- First split point
root directory
The top-level:
- Folder in filesystem
- Container for ML project files
- Reference point for paths
RMSE
Root Mean Squared Error:
- sqrt(mean(squared errors))
- Common regression metric
- Sensitive to outliers
rotational invariance
Property where:
- Input rotations don't affect output
- Important for image models
- Achieved via data augmentation
ROUGE
Family of metrics for:
- Evaluating text summarization
- Comparing machine/human text
- Measuring n-gram overlap
Variants:
- ROUGE-N (n-gram based)
- ROUGE-L (longest common subsequence)
- ROUGE-S (skip-bigram)
R-squared
Coefficient of determination:
- Measures variance explained
- Ranges 0-1 (higher is better)
- Common regression metric