Learn from Anywhere: Rethinking Generalized Zero-Shot Learning with Limited Supervision

TL;DR

This paper introduces a practical inductive zero-shot and few-shot learning framework that leverages unlabeled out-of-data class samples to improve generalization, addressing limitations of existing transductive methods.

Contribution

It proposes a novel AUD module and a product-of-experts formulation to utilize out-of-data unlabeled samples, enhancing zero-shot learning without requiring data from target unseen classes.

Findings

Improved generalization in zero-shot and few-shot learning scenarios.

Effective use of out-of-data unlabeled samples with minimal annotation cost.

Applicability to generalized zero-shot learning with limited supervision.

Abstract

A common problem with most zero and few-shot learning approaches is they suffer from bias towards seen classes resulting in sub-optimal performance. Existing efforts aim to utilize unlabeled images from unseen classes (i.e transductive zero-shot) during training to enable generalization. However, this limits their use in practical scenarios where data from target unseen classes is unavailable or infeasible to collect. In this work, we present a practical setting of inductive zero and few-shot learning, where unlabeled images from other out-of-data classes, that do not belong to seen or unseen categories, can be used to improve generalization in any-shot learning. We leverage a formulation based on product-of-experts and introduce a new AUD module that enables us to use unlabeled samples from out-of-data classes which are usually easily available and practically entail no annotation cost. In addition, we also demonstrate the applicability of our model to address a more practical and challenging, Generalized Zero-shot under a limited supervision setting, where even base seen classes do not have sufficient annotated samples.