Saved in:
| Main Authors: | , |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2025
|
| Online Access: | https://doi.org/10.5281/zenodo.17823481 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- Extreme few-shot learning (EFL) presents a significant challenge in machine learning, where models must generalize from an extremely limited number of examples, often just one or two. Current state-of-the-art approaches primarily rely on meta-learning strategies that optimize for rapid adaptation by leveraging prior experience across numerous tasks. However, many of these methods remain heavily dependent on the direct processing and comparison of specific exemplars, whether through learned feature embeddings, metric-based comparisons, or gradient-based fine-tuning. This reliance can limit their ability to generalize effectively to entirely novel classes or domains that deviate significantly from the distribution of observed exemplars. This paper proposes a conceptual shift in EFL, moving beyond a sole focus on exemplar-based adaptation towards learning and leveraging more fundamental, generalizable abstractions. We argue that true extreme few-shot generalization necessitates the extraction of underlying concepts, causal relationships, or structural invariants that transcend specific instances. We explore theoretical foundations for this paradigm shift, review existing methods and their limitations, and outline a methodological framework for developing models capable of distilling abstract knowledge. The paper discusses the potential for hybrid approaches combining neural network representational power with symbolic reasoning or explicit knowledge graphs, aiming to bridge the gap between perceptual data and conceptual understanding. Ultimately, we envision a future for EFL where models can synthesize new knowledge not just by reconfiguring existing features, but by understanding the abstract principles governing those features and their relationships, leading to more robust, interpretable, and human-like intelligence in data-scarce scenarios.