New AI System Learns Interpretable Image-Labeling Rules Without Human Feedback

Accurate image labeling underpins reliable computer vision systems, but building large-scale labeled datasets remains expensive and labor-intensive, especially in specialized domains that require expert knowledge. A newly published study in Applied Intelligence proposes a way to automate this process end to end while keeping the results interpretable to humans.

The method, called Iterative Dual-Branch First-Order Inductive Learning (IDB-FOIL), was developed by Yuwei Deng and Hong Wu and described in an article published on 27 October 2026 in the journal Applied Intelligence (Volume 56, article number 520). It combines the perceptual strength of pre-trained vision models with the transparency of rule-based reasoning, a combination known as neuro-symbolic learning.

According to the authors, existing rule-based approaches to image labeling still depend on human feedback to refine the rules they learn, and they tend to favor attributes that appear frequently in the data. Both limitations restrict scalability and reduce the diversity of the learned rules. IDB-FOIL is designed to eliminate manual rule refinement entirely.

Technically, the algorithm runs two parallel FOIL branches, referring to the classical First-Order Inductive Learning procedure for learning logical rules. One branch is guided by the original information-based metric, which favors frequent predicates. The other uses a frequency-independent metric that treats frequent and infrequent predicates equally, allowing rare but informative attributes to contribute to the learned rules.

The two branches operate iteratively with decayed predicate selection metrics. This decay progressively penalizes predicates that are selected often or that can be replaced by others, pushing the system toward rules that are diverse and complementary rather than redundant.

When the system labels an image, it uses a tf–idf-weighted voting mechanism to resolve labeling conflicts and improve the reliability of its decisions.

The researchers evaluated IDB-FOIL on four benchmark datasets drawn from both specialized and general domains. They report that the method achieved high labeling accuracy, high rule diversity, and strong interpretability, while fully removing the need for manual rule refinement. In their words, the results show that IDB-FOIL provides a scalable, interpretable, and fully automated solution for image labeling.

The source code of the research has been made publicly available on GitHub, allowing other researchers to examine and build on the approach.

Subject of Research: A neuro-symbolic inductive logic learning algorithm for automated, interpretable image labeling

Article Title: IDB-FOIL: iterative dual-branch inductive logic learning for automated and interpretable image labeling

Article References: Deng, Y., & Wu, H. (2026). IDB-FOIL: iterative dual-branch inductive logic learning for automated and interpretable image labeling. Applied Intelligence, 56(16), 520. https://doi.org/10.1007/s10489-026-07394-9

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07394-9

Keywords: neuro-symbolic learning, inductive logic programming, image labeling, computer vision, interpretable AI, rule-based reasoning, FOIL, machine learning, automated annotation, tf-idf voting, IDB-FOIL, iterative