Exploring Latent Class Structures in Classification- By-Components Networks Lars Holdijk Webcam image missing www.larsholdijk.com Radboud University Nijmegen larsholdijk larsholdijk@gmail.com larsholdijk@gmail.com
Example: Two classes in the ImageNet dataset Describe relation between object classes Observation: Webcam image missing Both objects share a number of components; wheels, doors, windshield, etc. • Some components are unique to the objects; raised roof and spoiler • www.larsholdijk.com larsholdijk Discriminative features larsholdijk@gmail.com Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk
Latent Class Structure “The information about the relationship between classes encoded in the way they share components.” Observation: Webcam image missing Investigating if, and how, classifiers use the latent class structure in their • classification process can provide useful information. www.larsholdijk.com larsholdijk larsholdijk@gmail.com Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk
Classification-By-Components Network Saralajew et al, Classification-by-Components: Probabilistic Modeling of Reasoning over a Set of Components, NeurIPS2019 � � � � � � � � Webcam image missing Shared Component Graph: 2 Nodes represent classes • www.larsholdijk.com Classes sharing a component are connected larsholdijk • 1 3 with an edge larsholdijk@gmail.com Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk
eNet and CUB bird species Evaluation: Im Imag ageN Great Dane hyena paper towel candle giant panda custard apple snow koala (f) (a) leopard leopard Pekinese French bulldog dalmatian barn (c) matchstick (e) (d) english setter jaguar cheetah street car pug trolley bus tiger cat tiger (b) boathouse Observation: Webcam image missing CBCs use the latent class structure in its classification process. • eg: large clusters of animals • www.larsholdijk.com larsholdijk larsholdijk@gmail.com Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk
eNet and CUB bird species Evaluation: Im Imag ageN Observation: Webcam image missing Consider WordNet hierachy as ground-truth for class relation. • Classes close to eachother in the WordNet hierachy more often share a • www.larsholdijk.com larsholdijk component than classes in completely different subtrees. larsholdijk@gmail.com Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk
Evaluation: ImageNet and CU CUB B bird species Pilated Red headed Belted Woodpecker woodpecker Kingfisher Yellow breasted chat Cardinal American Green tailed Pine Redstart Towhee Grosbeak Yellow headed Blackbird Gray Kingbird Anna Hummingbird American Bewick Goldfinch Wren White necked Observation: Raven Webcam image missing CBCs can fail to uncover the latent class structure when overparameterized. • With an ample number of components, specialized class specific components • www.larsholdijk.com larsholdijk can be formed. larsholdijk@gmail.com Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk
Exploring Latent Class Structures in Classification-By- Components Networks Webcam image missing www.larsholdijk.com larsholdijk larsholdijk@gmail.com
Recommend
More recommend