AI 용어집
인공지능 완전 사전
Interpretable AutoML
Subfield of AutoML that aims to automatically generate machine learning models that simultaneously optimize predictive performance and human interpretability of decisions.
Performance-Interpretability Trade-off
Fundamental dilemma in AutoML where improving predictive accuracy often comes with decreased model interpretability, requiring an optimal balance depending on the use case.
Local Feature Importance
Evaluation of the impact of each feature on a specific prediction, explaining why the model made a particular decision for a given individual.
Explainability vs Interpretability
Distinction where interpretability concerns intrinsic understanding of the model mechanism, while explainability aims to provide understandable justifications of predictions, even for opaque models.
Surrogate Models
Simple and interpretable models trained to approximate the behavior of a complex model, allowing generation of global or local explanations of the original model's decisions.
Model-agnostic Explanations
Interpretation techniques that can be applied to any type of machine learning model without requiring knowledge of its internal architecture.
Rule-based Models
Models using sets of interpretable IF-THEN rules to make decisions, offering a natural balance between performance and transparency in AutoML.
Post-hoc Interpretability
Approach consisting of analyzing and explaining an already trained model without modifying its structure, unlike intrinsically interpretable models.
Automated Feature Selection
AutoML process that automatically identifies the optimal subset of predictive variables that maximizes performance while preserving the interpretability of the final model.
Fidelity Score
Metric that evaluates the fidelity of an explanation or surrogate model by measuring how faithfully it reproduces the predictions of the original model.
Interpretable Neural Networks
Neural network architectures specifically designed to be interpretable, such as networks with monotonicity constraints or prototype networks.