AI 용어집
인공지능 완전 사전
Intrinsic Explanation
Approach where interpretability is directly integrated into the model's structure from its design, making the model naturally transparent.
Black Box Model
AI system whose internal workings are opaque or complex, making it difficult to directly understand its decision-making mechanisms.
White Box Model
Intrinsically transparent model whose internal mechanisms and decision processes are directly observable and understandable.
Sensitivity Analysis
Post-hoc technique evaluating how variations in input features affect the model's predictions to identify influential factors.
Naturally Interpretable Model
Intrinsic architecture designed to be transparent, such as decision trees or linear regression, not requiring external explanations.
Local Approximation
Post-hoc approach explaining a specific prediction by creating a simple model that approximates the complex model's behavior only in the neighborhood of that prediction.
Global Approximation
Post-hoc method aiming to create a simplified model that mimics the general behavior of the complex model across its entire input space.
Perturbation Method
Post-hoc technique analyzing the impact on predictions by systematically modifying input features to understand their role in the decision.
Saliency Map
Post-hoc visualization showing the most influential regions or features in the input data for a specific model prediction.
Interpretable Decision Tree
Intrinsic model using a hierarchical structure of if-then rules to make decisions, offering complete traceability of reasoning.
Counterfactual Analysis
Post-hoc method identifying the minimal changes needed in input features to change the model's prediction to a different outcome.
Proxy Model
Simple and interpretable model trained post-hoc to mimic the behavior of a complex model, serving as an approximation to facilitate interpretation.
Interpretable Hybrid Model
Intrinsic architecture combining complex components with built-in interpretation mechanisms to balance performance and transparency.
Feature Attribution
Post-hoc process assigning contribution scores to each input feature to explain their individual role in a specific prediction.