Learning Classifiers for Target Domain with Limited or No Labels - - PowerPoint PPT Presentation

Boston University Data Science & Machine Learning Lab

Learning Classifiers for Target Domain with Limited or No Labels

Pengkai Zhu, Hanxiao Wang, Venkatesh Saligrama

Boston University Slideshow Title Goes Here Boston University Data Science & Machine Learning Lab

Resource-limited Classification

Learning Classifiers for Target Domain with Limited or No Labels

06/12/2019 Wed

Task Target Domain What’s new? Example? Label? Domain Adaptation input Yes No Few-Shot Learning class Few Few Zero-Shot Learning class No No

“train from scratch” is impossible → Adapt existing models to new environment ✔ Goal: A universal, static representation robust to domain shift

Boston University Slideshow Title Goes Here Boston University Data Science & Machine Learning Lab

Low-Dimensional Visual Attributes (LDVA) Encoding

Learning Classifiers for Target Domain with Limited or No Labels

06/12/2019 Wed

Pa Part Att ttenti tion Mod

• del

High-dim Visual Feature

Pa Part Fea Feature Ex Extr tracto tor

Boston University Slideshow Title Goes Here Boston University Data Science & Machine Learning Lab

LDVA Train

Learning Classifiers for Target Domain with Limited or No Labels

06/12/2019 Wed

High-dim Visual Feature

𝜌 𝑙|𝑛

𝝆𝒏,𝒍: Probability of part 𝒏 belongs to type 𝒍

LDV LDVA En Encoding

Pa Part Feat Feature En Encoder

Type-1

×0.79 + ×0.03 + ×0.01+ ⋯

Type-2 Type-3

Part Fea eature Deco Decoder

Type-1

×0.01 + ×0.84 + ×0.03+ ⋯

Type-2 Type-3

Boston University Slideshow Title Goes Here Boston University Data Science & Machine Learning Lab

LDVA - Inference

Learning Classifiers for Target Domain with Limited or No Labels

06/12/2019 Wed

High-dim Visual Feature

𝜌 𝑙|𝑛

𝝆𝒏,𝒍: Probability of part 𝒏 belongs to type 𝒍

LDV LDVA En Encoding

Pa Part Feat Feature En Encoder Semantic Attributes Eye color: black Crown color: blue Wing color: green Breast color: red …

Neare rest t Neig ighbor Cl Classificat ation

Generalized Zero-Shot Learning Domain Adaptation Few-Shot Learning

Boston University Slideshow Title Goes Here Boston University Data Science & Machine Learning Lab

Comparison with other methods

Learning Classifiers for Target Domain with Limited or No Labels

06/12/2019 Wed

▪ Vanilla DNN: ▪ Attention Methods: ▪ Ours:

NN

Input high-dim feature

NN

Input attention High-dim feature attention High-dim feature

NN

Input attention Part Encoder attention Part Encoder Low-dim LDVA Low-dim LDVA

Boston University Slideshow Title Goes Here Boston University Data Science & Machine Learning Lab

Low-Dimensional Visual Attributes (LDVA) Encoding ▪ Every object is encoded into a mixture of part types ▪ Benefits:

▪ Low-dimensional: proto-types in each part is limited ▪ Compositional Uniqueness: every class is represented uniquely ▪ Small intra-class variance and large inter-class variance ▪ Robust to domain shift

Learning Classifiers for Target Domain with Limited or No Labels

06/12/2019 Wed

Boston University Slideshow Title Goes Here Boston University Data Science & Machine Learning Lab

Low-Dimensional Visual Attributes (LDVA) Encoding ▪ Every object is encoded into a mixture of part types ▪ Benefits:

▪ Low-dimensional: proto-types in each part is limited ▪ Compositional Uniqueness: every class is represented uniquely ▪ Small intra-class variance and large inter-class variance ▪ Robust to domain shift ▪ Mirrors human-labeled semantic vector ▪ Encode unseen class by seen part-types ▪ Requires less data and feedback

Learning Classifiers for Target Domain with Limited or No Labels

06/12/2019 Wed

Boston University Slideshow Title Goes Here Boston University Data Science & Machine Learning Lab

Experiments

Learning Classifiers for Target Domain with Limited or No Labels

06/12/2019 Wed

Generalized Zero-Shot Learning

Boston University Slideshow Title Goes Here Boston University Data Science & Machine Learning Lab

Experiments

▪ Few-Shot Learning ▪ Domain Adaptation

Learning Classifiers for Target Domain with Limited or No Labels

06/12/2019 Wed

Boston University Data Science & Machine Learning Lab

Welcome to our poster today! Thank you!

#133

Pacific Ballroom