Yassir Bendou Thesis defense

Few-shot learning, which focuses on effective adaptation with minimal labeled examples, is vital in scenarios where data is scarce, expensive, or challenging to acquire. This thesis explores how different data modeling approaches in few-shot image classification can enhance a model's capacity to generalize and adapt to novel, unseen tasks.  By bringing a data modeling perspective to address the bias-variance trade-off, this work explores how setting assumptions on data distributions can guide the design of solutions that balance expressivity and generalization. In addition to advancing single-modal methods, the thesis delves into multi-modal few-shot learning, leveraging the textual descriptions to improve cross-modal generalization. Contributions include developing a robust Gaussian-based baseline for few-shot classification, utilizing kernel methods to estimate Bayes-optimal solutions in a multi-modal context, and introducing methods for predicting generalization performance when validation data is unavailable. The work also introduces a novel problem, zero-shot one-class classification, which discriminates categories based on label descriptions alone. Through a combination of theoretical insights and practical benchmarks, this thesis demonstrates how thoughtful data modeling choices can effectively address key challenges in few-shot learning.