Reproduce “Transductive One-Shot Learning Meet Subspace Decomposition” on OpenResearch

About this paper

Kyle Stein, Andrew A. Mahyari, Guillermo Francia III, Eman El-Sheikh

A transductive one-shot learning method, developed at the University of West Florida, leverages subspace decomposition to identify fundamental image primitives from embeddings for robust classification of novel classes. This approach achieves higher accuracy on mini-ImageNet and tiered-ImageNet in several 5-way and 10-way 1-shot classification tasks compared to existing state-of-the-art methods.

Key insights

• Decomposing image embeddings into latent primitives through subspace decomposition provides a more structured and interpretable feature representation, which enhances generalization to novel classes by uncovering underlying compositional features.
• Explicitly modeling these latent variables (referred to as "primitives" or "basis vectors") and their combinations can more effectively group similar samples and propagate labels than direct feature similarity metrics alone.
• The concurrent learning of basis and coefficient matrices during the decomposition process, coupled with specific non-negativity and categorical constraints, contributes to the model's stability and robust performance in low-data settings.

Summary by alphaXiv.