AI 용어집
인공지능 완전 사전
ICE Plot
Visualization showing how the prediction changes for each individual observation when a feature varies, unlike PDP which shows the average effect. ICE plots reveal heterogeneity of effects and subgroups of behavior.
ALE Plot
Alternative to PDP that calculates local effects by averaging prediction differences in feature space regions. ALE plots avoid biases of PDP when features are correlated.
Activation Maps
Heatmaps overlaid on input images showing regions that most strongly activate neurons in convolutional neural networks. Activation maps visually explain what the model detected in an image.
t-SNE Visualization
Non-linear projection of high-dimensional data to 2D or 3D preserving local relationships between points to reveal clusters. t-SNE helps understand the latent structure of representations learned by the model.
UMAP Projection
Dimensionality reduction technique alternative to t-SNE offering better preservation of global structure and faster execution speed. UMAP reveals underlying manifolds in embeddings of complex models.
Model Comparison Radar Chart
Spider web visualization simultaneously comparing multiple models according to different performance metrics. This chart facilitates multi-criteria evaluation and selection of the best model according to priorities.
Permutation Importance Visualization
Graph showing the performance drop of the model when values of a feature are randomly permuted, thus evaluating its real importance. This method captures non-linear effects and interactions better than classical importance.