Infomation Ricoh paper on reliable AI development with limited data accepted for poster presentation at IJCNN 2026 Evaluating the practicality of AI that can recognize when it is unable to predict a reliable answer using existing multimodal foundation models

TOKYO, June 19, 2026 – Ricoh announced today that a paper on its technology for developing reliable AI models with limited data has been accepted for poster presentation at the International Joint Conference on Neural Networks (IJCNN) 2026, one of the leading international conferences in the field of neural networks, a foundation technology for artificial intelligence (AI). The technology enables AI systems to recognize when they are unable to predict a reliable answer, helping improve the reliability of AI in practical applications with limited training data.

Held in the Netherlands from June 21 to 26, 2026, the conference is jointly sponsored by the Institute of Electrical and Electronics Engineers (IEEE) and the International Neural Network Society (INNS).

While AI adoption continues to expand across a wide range of industries, obtaining sufficient training data remains a challenge. In addition, AI systems may sometimes behave as if they know the correct answer even when presented with unfamiliar or previously unseen data, raising concerns about the reliability of AI decision-making. In practical applications, there is a growing need for AI that can not only make accurate decisions with limited data, but also recognize when it does not know the answer.

To address these challenges, the paper proposes a method that combines a Bayesian machine learning model capable of evaluating prediction reliability with Contrastive Language-Image Pre-training (CLIP), a multimodal foundation model that captures relationships between images and text. By evaluating image-text similarity using separate criteria, the method quantitatively estimates predictive uncertainty. This enables AI systems to recognize that they are unable to predict a reliable answer when presented with previously unseen inputs. Furthermore, by employing a training-free optimization approach, the method reduces the need for additional training during deployment and enables rapid adoption across a wide range of practical applications.

The paper was selected based on its demonstrated practicality in improving performance using existing multimodal foundation models with minimal additional training. It was also recognized for its ability to handle previously unseen data and maintain stable performance under diverse conditions. The technology is expected to enhance the reliability of AI and expand its range of applications in areas where avoiding misjudgments is crucial, such as visual inspection in manufacturing and inspection of equipment and infrastructure.

Ricoh will continue to accelerate its research and development efforts by leveraging its expertise in neural networks and will further enhance the technologies needed to develop and deploy reliable AI rapidly, even in environments with limited training data. Through these efforts, Ricoh aims to deliver trustworthy AI services tailored to customers' industries and business needs, creating new value and supporting Fulfillment through Work.