Introduction
In computer vision, few-shot learning remains a significant challenge. While humans can generalize from a single example by leveraging a lifetime of background knowledge, deep learning models typically require vast datasets to achieve similar proficiency. The primary bottleneck is the "data scarcity" problem: when a model sees only a few examples of a new class, it lacks the context to generalize effectively. Most current methods rely on metric learning—calculating distances between image features in a vector space—but these features often lack inherent semantic structure.
A new framework called ViOCE (Vision with Ontology-based Concept Embeddings) addresses this by bridging the gap between symbolic logic and continuous geometric representations. By teaching models to "think" in hierarchies using $n$-ball embeddings, ViOCE allows AI to leverage structured background knowledge, such as WordNet, to guide visual training.
Methodology and Architecture
The core innovation of ViOCE is the translation of rigid, symbolic ontologies into a geometric format that neural networks can utilize. The process begins with Ontology Construction, where relationships like "SubClassOf" or "DisjointWith" are extracted from knowledge resources.
From Logic to Geometry
Instead of representing a concept (e.g., "Dog") as a single point, ViOCE represents it as an $n$-ball—a sphere in high-dimensional space defined by a center ($c$) and a radius ($r$). This volume-based approach enables two critical logical properties:
- Subsumption: If "Poodle" is a subclass of "Dog," the $n$-ball for Poodle is mathematically constrained to reside within the $n$-ball for Dog.
- Disjointness: If "Dog" and "Street Sign" are disjoint concepts, their respective $n$-balls are pushed apart in the embedding space.
To ensure logical consistency, the system computes an Inferred Class Hierarchy (ICH) to capture transitive relations. A standard backbone (ResNet50) extracts image features, which a Multi-Layer Perceptron (MLP) then maps into the same $n$-dimensional space as the concept balls. The objective is to project an image directly into its corresponding concept's $n$-ball.
Geometric Representation
Representing classes as volumes rather than points provides significant flexibility for intra-class variation. Two distinct-looking poodles can land in different parts of the same "Poodle" ball while still being correctly classified.
This structure also acts as a semantic "safety net." If the model fails to identify a specific breed (the "Poodle" ball), it may still correctly project the image into the parent "Dog" ball. This makes the model's errors more semantically meaningful and provides a layer of transparency often missing in black-box neural networks.
Key Results and Performance
ViOCE was evaluated on standard benchmarks, including miniImageNet and tieredImageNet, demonstrating state-of-the-art (SOTA) performance:
- miniImageNet (5-way, 5-shot): Achieved 93.65% accuracy, a substantial improvement over previous SOTA methods (~82%).
- 20-way Classification: In a more complex 20-way, 5-shot task, ViOCE reached 84.13% accuracy, significantly outperforming the next best model (TIM-GD) at 59.50%.
- Generalization: The model showed higher testing accuracy (95.36%) than training accuracy (85.32%) during base learning, suggesting that the $n$-ball structure forces the learning of highly generalizable features.
Conclusion
By transforming abstract logic into tangible geometry, ViOCE demonstrates that structured background knowledge is a vital component for efficient machine vision. While the system currently relies on high-quality ontologies like WordNet, future work aims to incorporate multi-relational knowledge (e.g., "part-of" relations) to further refine the geometric space and improve the model's handling of complex semantic hierarchies.
Summary updated