Exploring Latent Class Structures in Classification- By-Components - - PowerPoint PPT Presentation

▶

Aug 13, 2023 182 likes •278 views

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

SLIDE 1

Exploring Latent Class Structures in Classification- By-Components Networks

Lars Holdijk Radboud University Nijmegen larsholdijk@gmail.com

Webcam image missing

www.larsholdijk.com larsholdijk larsholdijk@gmail.com

SLIDE 2

Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk

Observation:

Both objects share a number of components; wheels, doors, windshield, etc.
Some components are unique to the objects; raised roof and spoiler

Describe relation between object classes

Example: Two classes in the ImageNet dataset

Webcam image missing

www.larsholdijk.com larsholdijk larsholdijk@gmail.com

Discriminative features

SLIDE 3

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:

Investigating if, and how, classifiers use the latent class structure in their

classification process can provide useful information. Webcam image missing

www.larsholdijk.com larsholdijk larsholdijk@gmail.com

SLIDE 4

Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk

Webcam image missing

www.larsholdijk.com larsholdijk larsholdijk@gmail.com

Classification-By-Components Network

Saralajew et al, Classification-by-Components: Probabilistic Modeling

f Reasoning over a Set of Components, NeurIPS2019

Shared Component Graph:

Nodes represent classes
Classes sharing a component are connected

with an edge 3 2 1

SLIDE 5

Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk

Webcam image missing

www.larsholdijk.com larsholdijk larsholdijk@gmail.com

Evaluation: Im Imag ageN eNet and CUB bird species

giant panda Great Dane dalmatian jaguar english setter snow leopard hyena koala leopard cheetah tiger tiger cat candle matchstick paper towel custard apple Pekinese French bulldog pug barn boathouse (a) (c) (d) street car trolley bus (b) (e) (f)

Observation:

CBCs use the latent class structure in its classification process.
eg: large clusters of animals

SLIDE 6

Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk

Webcam image missing

www.larsholdijk.com larsholdijk larsholdijk@gmail.com

Evaluation: Im Imag ageN eNet and CUB bird species

Observation:

Consider WordNet hierachy as ground-truth for class relation.
Classes close to eachother in the WordNet hierachy more often share a

component than classes in completely different subtrees.

SLIDE 7

Exploring Laten Class Structures in Classification-By-Components Networks Lars Holdijk

Webcam image missing

www.larsholdijk.com larsholdijk larsholdijk@gmail.com

Evaluation: ImageNet and CU CUB B bird species

Observation:

CBCs can fail to uncover the latent class structure when overparameterized.
With an ample number of components, specialized class specific components

can be formed.

Red headed woodpecker Pilated Woodpecker Gray Kingbird American Redstart Yellow breasted chat Belted Kingfisher Green tailed Towhee Bewick Wren White necked Raven American Goldfinch Cardinal Anna Hummingbird Pine Grosbeak Yellow headed Blackbird

SLIDE 8

Exploring Latent Class Structures in Classification-By- Components Networks

Webcam image missing