A Gating Model for Bias Calibration in Generalized Zero-Shot Learning

Personnel: Gykyeong Kwon

Goal:  Generalized zero-shot learning (GZSL) aims to train a model that can generalize to unseen class data by only using auxiliary information. 

Challenges: GZSL models are generally more biased towards seen classes due to the overfitting on the seen class data. 

Our Work: We utilize a two-stream autoencoder-based gating model for GZSL, where the gating model predicts whether the query data is from seen classes or unseen classes. With two separate experts for seen and unseen classes, the class predictions can be made independently from each other.

References:

1. G. Kwon and G. AlRegib, "A Gating Model for Bias Calibration in Generalized Zero-Shot Learning, " in IEEE Transactions on Image Processing (TIP), Feb. 10 2022.