Saralajew, Sascha: New Prototype Concepts in Classification Learning. 2020
Inhalt
- Introduction
- Learning Vector Quantization
- Generalized Tangent Learning Vector Quantization
- Motivation
- Set-prototypes and respective learning vector quantization variants
- Generalized tangent learning vector quantization: Affine subspace prototypes
- Restricted generalized tangent learning vector quantization: ns-orthotope prototypes
- Relations to other concepts
- Accuracy and interpretability evaluations
- Generalized tangent learning vector quantization as margin maximizer
- Related work
- Summary and discussion
- Classification-by-Components Networks
- Motivation
- Probabilistic modeling of reasoning over a set of components
- Reasoning over a set of full-size components
- Reasoning over a set of patch components
- Multiple components and reasoning strategies
- General remarks
- Joint training with a trainable feature extractor
- Evaluation without a feature extractor
- Evaluation with a feature extractor
- MNIST: Ablation study
- MNIST: Varying the number of components
- MNIST: Initial robustness and rejection evaluation
- MNIST: Interpretation of the reasoning process
- GTSRB
- CIFAR-10
- ImageNet
- Related work
- Summary and discussion
- Summary and Concluding Remarks
- Publications
- Mathematical Symbols
- Acronyms
- References