VisionDES: Deep Dynamic Ensemble Selection for Trustworthy Image Classification

Hyperparameters of VisionDES

Parameter	Description	Default
pool_classifiers	List of classifiers	None
k	Number of neighbors in RoC	7
return_logits	Whether to return raw logits instead of just prediction	False
explain	Enable explanation mechanism	False
top	Select only top n classifiers	False
n	Number of top classifiers to select	3
FIRE	Use FIRE-based filtering mechanism	False
Per-class minimum	Min number of classifiers per class in selection (FIRE)	1
similarity	Enable similarity-based weighting in aggregation	True
similarity_threshold	Minimum similarity for including a selected ensemble	0
alpha	Weighting factor for combining competence and similarity	0.6
knorae	Use KNORA-E instead of default KNORA-U for selection	False

Requirements

• Python >= 3.9
• PyTorch == 2.8.0
• Torchvision == 0.19.0

Example for VisionDES XAI