Newswise — Korean researchers have developed a core supply expertise that enables multimodal synthetic intelligence (AI) to reliably retain present data with out shedding it even when repeatedly studying new data, attracting international consideration
Electronics and Telecommunications Research Institute (ETRI) introduced that the “Continual and Compositional Knowledge Editing Technology (MemEIC),” collectively developed by the analysis crew led by Director Lim Soo Jong of the Language Intelligence Research Section with Pohang University of Science and Technology and Sungkyunkwan University, was offered at NeurIPS 2025, the world’s most prestigious AI convention, and held in San Diego, U.S., late final 12 months.
Recently, multimodal AI that may perceive photos and textual content concurrently, corresponding to ChatGPT, Gemini, and Claude, has been spreading quickly. For instance, it’s AI that may have a look at a photograph and describe it or reply when requested in textual content about what’s in a picture.
However, one of these AI had one main downside. When AI learns new info or revises present info, a phenomenon referred to as “catastrophic forgetting” happens, through which it forgets beforehand realized data as effectively. In different phrases, when AI learns one thing new, it develops a sort of “AI forgetfulness” through which it forgets what it beforehand knew.
In specific, when visible info and language info needed to be revised on the identical time, the 2 kinds of data usually grew to become combined collectively, inflicting the AI to misconceive and incessantly give improper solutions to compositional questions.
For instance, if an AI is sequentially taught the visible info, “The dessert shown in the photo is Dubai chewy cookie (Dujjonku),” and the language info, “Dujjonku is popular in Korea,” and is then requested, “In which country is this dessert popular?”, present AI fashions confirmed limitations in correctly connecting the picture with the related data.
In such circumstances, present fashions incessantly produced hallucinations, corresponding to misidentifying the dessert within the picture and producing inaccurate solutions corresponding to, “The image shown in the photo is a chocolate truffle and is popular in Europe.”
To remedy this downside, ETRI researchers developed a knowledge-editing AI expertise able to precisely answering even compositional questions.
Conventional strategies primarily used an method that modified data by straight modifying the AI’s inner parameters. This was a sort of “brain-surgery-like approach” that basically altered the construction of the prevailing mannequin, and it was restricted in that even beforehand saved info may very well be affected within the strategy of modifying data.
Instead, the researchers proposed a way of storing new info in exterior reminiscence relatively than contained in the AI. This method of including an auxiliary reminiscence has a construction that retrieves and makes use of info solely when wanted, making it attainable to flexibly add new info whereas sustaining the soundness of the prevailing mannequin, thereby guaranteeing scalability as effectively.
MemEIC was designed with inspiration from the construction of the human mind. Just because the human mind is split into the left and proper hemispheres, which play completely different roles, the AI was designed to retailer data individually as effectively.
Image-related visible info is saved independently in a “visual adapter,” whereas text-related language info is saved independently in a “language adapter.” Then, when the AI receives a compositional query that requires understanding each photos and textual content, a “knowledge connector” hyperlinks the 2 items of data in keeping with context to supply a solution.
It was confirmed that AI making use of the MemEIC expertise developed by ETRI precisely mixed visible and language info and appropriately answered, “The dessert shown in the photo is Dubai chewy cookie (Dujjonku), and it is popular in Korea.”
In this manner, by means of a separated-storage and selective-combination construction that shops data individually and connects it solely when wanted, an AI structure able to compositional reasoning and answering compositional questions was applied by minimizing the issues of inner interference, through which completely different info turns into combined, and the degradation of present data.
To confirm the expertise’s efficiency, the researchers constructed a compositional data modifying benchmark (CCKEB) consisting of 1,278 gadgets and performed experiments that sequentially edited a whole lot of items of information. As a outcome, MemEIC achieved an accuracy degree of roughly 70% in answering compositional questions. Compared with the 36% to 52% vary of present applied sciences, this represents greater than double the efficiency. In addition, the preservation attribute of “locality,” through which response stability is maintained as a result of solutions to present questions don’t change even after new data is added, was additionally confirmed.
This research is very significant in that it goes a step past merely assuaging AI’s forgetting phenomenon and concurrently solves the 2 challenges of steady data modifying and compositional reasoning. In specific, it’s anticipated to have excessive sensible applicability in clever service fields that require continuous updates as new info is continually added and altered, corresponding to coverage and authorized info, product info, and industrial information.
Director Lim Soo Jong of ETRI’s Language Intelligence Research Section mentioned, “This study is an achievement that has laid the technological foundation for multimodal AI to simultaneously reflect up-to-date information required in real service environments and secure reliability. We will further advance the technology so that it can reliably reflect diverse information from industrial sites in the future.”
Seong Jin, the primary creator of the paper and a researcher at ETRI’s Language Intelligence Research Section, defined, “Conventional methods had the problem of interference occurring while revising visual knowledge and language knowledge at the same time. MemEIC overcame these limitations through a structure that stores the two types of knowledge independently and connects them only when needed.”
###
This research was performed as a part of the “Development of Learning and Utilization Technologies for the Sustainability of Generative Language Models and the Reflection of Up-to-Dateness Over Time” undertaking below the “Next-Generation Generative AI Technology Development Program” supported by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation (IITP).
About Electronics and Telecommunications Research Institute (ETRI)
ETRI is a non-profit government-funded analysis institute. Since its basis in 1976, ETRI, a world ICT analysis institute, has been making its immense effort to offer Korea a exceptional development within the subject of ICT trade. ETRI delivers Korea as one of many high ICT nations within the World, by unceasingly growing world’s first and greatest applied sciences.