A paper from NTT laboratories has been accepted for publication for Neurocomputing

A paper submitted by NTT laboratories has been accepted to Neurocomputing, a highly regarded international journal in the field of neural networks and artificial intelligence (Impact Factor: 6.5).

The abbreviated names of the research institutes listed in the affiliations correspond to the following:

• CD: Computer and Data Science Laboratories
• SI: Social Informatics Laboratories
• CS: Communication Science Laboratories

Atsutoshi Kumagai (CD/ SI), Tomoharu Iwata (CS), Taishi Nishiyama (SI), Yasutoshi Ida (CD), Yasuhiro Fujiwara (CS)

Various methods have been proposed to improve the efficiency of annotation required for training AI models. However, existing approaches suffer from degraded predictive performance when the number of data samples or annotators is limited. In this study, we propose a method that integrates a probabilistic model that treats annotator ability and true labels as latent variables with neural networks and acquires neural network embedding representations through meta-learning across multiple tasks. Through experiments, we confirmed that the proposed approach can effectively learn high-performance AI models from limited data via probabilistic inference in the embedding space. This achievement is expected to further expand the range of AI applications across various domains, including healthcare and security, particularly for tasks where obtaining a large amount of labeled data is challenging, such as rare diseases or newly emerging malware.