Concepts de base
A fuzzy extension of the Scene Identification and Tagging (SIT) algorithm that bootstraps structured and robust scene representations in a fuzzy ontology, enabling incremental learning and classification of scenes.
Résumé
The paper presents a fuzzy extension of the Scene Identification and Tagging (SIT) algorithm, which aims to address the limitations of the crisp implementation of SIT. The key points are:
The crisp SIT algorithm bootstraps structured knowledge representations in an OWL ontology, allowing for incremental learning and classification of scenes. However, it suffers from robustness issues when dealing with noisy sensory data.
The fuzzy SIT extends the crisp version by representing scene elements and relations using fuzzy degrees, and encoding cardinality restrictions as fuzzy sets. This allows for a more robust handling of noisy and vague inputs.
The fuzzy SIT preserves the key properties of the crisp SIT, such as incremental learning, structured representations, and intelligible classification. However, the fuzzy representations lead to less intuitive knowledge models compared to the crisp version.
The paper details the theoretical foundations of the fuzzy SIT, including the encoding of scenes as fuzzy beliefs, the learning of fuzzy scene categories, the structuring of a fuzzy memory graph, and the fuzzy classification of new scenes.
Experimental results show that the fuzzy SIT improves robustness to noisy inputs while maintaining consistency with the crisp SIT behavior. However, the fuzzy representations lead to less intelligible knowledge models.
The paper discusses the trade-offs between the robustness and intelligibility of the bootstrapped representations, and outlines future work to address the intelligibility limitations of the fuzzy SIT.
Stats
"At least 0.8 cups are in front of glasses."
"At least 1.2 cups are in front of glasses, at least 0.1 glasses are in front of glasses, and at least 0.7 cups are in front of cups."
Citations
"Fuzzy SIT is robust, preserves the properties of its crisp formulation, and enhances the bootstrapped representations. On the contrary, the fuzzy implementation of SIT leads to less intelligible knowledge representations than the one bootstrapped in the crisp domain."