Une perspective de modélisation des données sur la classification d’images avec peu d’exemples

Few-shot learning, which adapts from minimal labeled data, is crucial when data is scarce, expensive, or hard to obtain. This thesis explores how different data modeling approaches in few-shot image classification enhance a model’s ability to generalize to novel tasks. By addressing the bias-variance trade-off through assumptions on data distributions, it designs solutions that balance expressivity and generalization. Beyond advancing single-modal methods, the thesis delves into multi-modal few shot-learning, leveraging textual descriptions to improve cross-modal generalization. Key contributions include developing a robust Gaussian based baseline, utilizing kernel methods to estimate Bayes-optimal solutions in a multi-modal context, and introducing methods for predicting generalization performance without validation data. The work also introduces a novel problem, zero-shot one-class classification, which discriminates a single category based solely on its label description. Through theoretical insights and practical benchmarks, this thesis demonstrates how thoughtful data modeling choices can effectively address key challenges in few-shot learning.