Shared Latent Models for Zero-Shot Learning

We develop a framework for zero-shot learning (ZSL) with a goal towards understanding recognition for scenarios where certain class-specific training data could be limited or unavailable. We explore ZSL as a prototypical instance of multimedia recognition to understand how multi-domain data (e.g. words, phrases, attributes) can help overcome lack of training data in the target domain (e.g. videos, images). We develop ZSL and its variations to investigate the value of having different degrees of information about test data. We pose ZSL as a class-independent binary hypothesis testing problem and develop shared latent representations for fusing multimedia data. We develop a joint discriminative learning framework to train class-independent latent representations and show that our resulting decision function learns to recognize similarity between source-target domain data pairs. We develop a modular optimization approach to account for various types of partial information that may become available during test time based on fusing predictions from learned similarity functions with any available test-time information. We test our method on several benchmark datasets, achieving state-of-the-art results.

Date:
Speakers:
Ziming Zhang
Affiliation:
Boston University

Watch Next