Reproduce “A statistical approach to bias in zero-shot learning: the lens of handwriting recognition” on OpenResearch

About this paper

Clarence Chew, Gim Siang Chia, Sukalpa Chanda, Subhroshekhar Ghosh, Soumendu Sundar Mukherjee

A statistical bias-correction layer is introduced for generalized zero-shot handwritten-word recognition. By first estimating whether an embedding belongs to a seen or unseen class and then restricting word matching accordingly, the method raises unseen-word accuracy from 0.770 to 0.931 and the harmonic mean from 0.840 to 0.917 on IAM, while reducing the original 769-dimensional embeddings to as few as 15–30 dimensions with much of the benefit retained.

Key insights

• Seen/unseen discrimination can be treated as a separate out-of-distribution classification problem before label selection, rather than as an intrinsic property of the feature learner.
• A stricter-than-majority ensemble threshold is useful because the underlying recognition model is systematically biased toward seen classes; ensemble size improves reliability under the paper’s theoretical assumptions.
• The embedding is highly structured: the first 15 and 30 principal components explain 62.24% and 75.92% of covariance, respectively, while preserving much of the bias-correction performance.

Summary by alphaXiv.