Learning Classifiers for Target Domain with Limited or No Labels Pengkai Zhu, Hanxiao Wang, Venkatesh Saligrama Boston University Data Science & Machine Learning Lab
Learning Classifiers for Target Domain with Limited or No Labels 06/12/2019 Wed Resource-limited Classification Boston University Slideshow Title Goes Here Target Domain Task 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 Data Science & Machine Learning Lab
Learning Classifiers for Target Domain with Limited or No Labels 06/12/2019 Wed Low-Dimensional Visual Attributes (LDVA) Encoding Boston University Slideshow Title Goes Here High-dim Visual Feature Pa Part Att ttenti tion Part Fea Pa Feature Mod odel Extr Ex tracto tor Boston University Data Science & Machine Learning Lab
Learning Classifiers for Target Domain with Limited or No Labels 06/12/2019 Wed LDVA Train Boston University Slideshow Title Goes Here High-dim Visual Feature Part Fea eature Deco Decoder Type-1 Type-2 Type-3 × 0.79 + × 0.03 + × 0.01+ ⋯ Type-1 Type-2 Type-3 Part Feat Pa Feature En Encoder × 0.01 + × 0.84 + × 0.03+ ⋯ LDVA En LDV Encoding 𝜌 𝑙|𝑛 𝝆 𝒏,𝒍 : Probability of part 𝒏 belongs to type 𝒍 Boston University Data Science & Machine Learning Lab
Learning Classifiers for Target Domain with Limited or No Labels 06/12/2019 Wed LDVA - Inference Boston University Slideshow Title Goes Here High-dim Visual Feature Semantic Attributes Eye color: black Generalized Zero-Shot Learning Crown color: blue Wing color: green Breast color: red … Pa Part Feat Feature Few-Shot Learning En Encoder Neare rest t Neig ighbor LDVA En LDV Encoding Domain Adaptation Cl Classificat ation 𝜌 𝑙|𝑛 𝝆 𝒏,𝒍 : Probability of part 𝒏 belongs to type 𝒍 Boston University Data Science & Machine Learning Lab
Learning Classifiers for Target Domain with Limited or No Labels 06/12/2019 Wed Comparison with other methods Boston University Slideshow Title Goes Here ▪ Vanilla DNN: NN Input high-dim feature ▪ Attention Methods: attention High-dim feature NN Input attention High-dim feature ▪ Ours: attention Part Encoder Low-dim LDVA NN Input attention Part Encoder Low-dim LDVA Boston University Data Science & Machine Learning Lab
Learning Classifiers for Target Domain with Limited or No Labels 06/12/2019 Wed Low-Dimensional Visual Attributes (LDVA) Encoding Boston University Slideshow Title Goes Here ▪ 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 Boston University Data Science & Machine Learning Lab
Learning Classifiers for Target Domain with Limited or No Labels 06/12/2019 Wed Low-Dimensional Visual Attributes (LDVA) Encoding Boston University Slideshow Title Goes Here ▪ 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 Boston University Data Science & Machine Learning Lab
Learning Classifiers for Target Domain with Limited or No Labels 06/12/2019 Wed Experiments Boston University Slideshow Title Goes Here Generalized Zero-Shot Learning Boston University Data Science & Machine Learning Lab
Learning Classifiers for Target Domain with Limited or No Labels 06/12/2019 Wed Experiments Boston University Slideshow Title Goes Here ▪ ▪ Few-Shot Learning Domain Adaptation Boston University Data Science & Machine Learning Lab
Thank you! #133 Welcome to our poster today! Pacific Ballroom Boston University Data Science & Machine Learning Lab
Recommend
More recommend