Ç¥ÁØÈ­ Âü¿©¾È³»

TTAÀÇ Ç¥ÁØÇöȲ

Ȩ > Ç¥ÁØÈ­ °³¿ä > TTAÀÇ Ç¥ÁØÇöȲ

Ç¥ÁعøÈ£ TTAK.KO-10.1594 ±¸Ç¥ÁعøÈ£
Á¦°³Á¤ÀÏ 2025-12-05 ÃÑÆäÀÌÁö 27
ÇѱÛÇ¥Áظí AI ¸ðµ¨ÀÇ ÇнÀ µ¥ÀÌÅͼ¼Æ® °ü¸®¿Í »óÈ£¿î¿ë¼º Áö¿øÀ» À§ÇÑ ¸ÞŸµ¥ÀÌÅÍ
¿µ¹®Ç¥Áظí Metadata for Supporting the Management and Interoperability of Training Datasets for AI Models
Çѱ۳»¿ë¿ä¾à ÀÌ Ç¥ÁØÀº ´Ù¾çÇÑ ÀΰøÁö´É(AI) ¸ðµ¨°ú °ü·Ã ÀÚ»ê(µ¥ÀÌÅͼÂ, ¹®¼­, ¼Ò½ºÄÚµå µî)ÀÇ ¼ö¸í ÁÖ±â Àü¹Ý¿¡ °ÉÄ£ ¸ÞŸµ¥ÀÌÅ͸¦ ±¸Á¶ÀûÀ¸·Î ±â¼úÇϱâ À§ÇÑ Ç¥ÁØ ¿ÂÅç·ÎÁö ½ºÅ°¸¶¸¦ Á¤ÀÇÇÏ¿´´Ù. AI ¸ðµ¨ÀÌ Á¡Â÷ º¹ÀâÇØÁö°í ÆÄ»ý ¸ðµ¨ÀÌ ºü¸£°Ô Áõ°¡ÇÔ¿¡ µû¶ó, °¢ ¸ðµ¨ÀÇ Ãâó(provenance), ÀçÇö¼º(reproducibility), ½Å·Ú¼ºÀ» È®º¸ÇÏ´Â °ÍÀÌ »ê¾÷ »ýŰèÀÇ Áö¼Ó°¡´É¼ºÀ» À§ÇØ ¹Ýµå½Ã ÇÊ¿äÇÏ´Ù. ƯÈ÷ ¸ðµ¨ÀÇ ¾ÆÅ°ÅØÃ³, ÇнÀ ¹× Æò°¡ °úÁ¤, Ȱ¿ëµÈ µ¥ÀÌÅͼ¼Æ®¿Í °°Àº ÇÙ½É Á¤º¸°¡ Ç÷§Æûº°·Î ÆÄÆíÈ­µÇ¾î Àְųª ºñÁ¤Çü ÅØ½ºÆ®·Î¸¸ Á¸ÀçÇÒ °æ¿ì, ¸ðµ¨ °£ÀÇ °´°üÀûÀÎ ¼º´É ºñ±³³ª ƯÁ¤ Á¶°ÇÀÇ ¸ðµ¨À» °Ë»öÇÏ´Â °ÍÀÌ ºÒ°¡´É¿¡ °¡±õ´Ù. ÀÌ·¯ÇÑ ¹®Á¦¸¦ ÇØ°áÇϱâ À§ÇØ, ÀÌ Ç¥ÁØÀº AI ¸ðµ¨ÀÇ ±â¼úÀû »ç¾ç»Ó¸¸ ¾Æ´Ï¶ó ÇнÀ, °ËÁõ, Æò°¡ µî Àüü ÇÁ·Î¼¼½º¿Í À±¸®Àû °í·Á»çÇ×±îÁö Æ÷°ýÇÏ´Â ÅëÇÕ ¸ÞŸµ¥ÀÌÅÍ ±¸Á¶¸¦ Á¦¾ÈÇÑ´Ù. AIModel, Dataset, TrainingProcess, ModelCard µîÀÇ ÇÙ½É Å¬·¡½º¿Í Ç¥ÁØ ¾îÈÖ(DCAT, SKOS, Schema.org) ±â¹ÝÀÇ ¼Ó¼ºµéÀ» ÅëÇØ, ´Ù¾çÇÑ ÃâóÀÇ ¸ðµ¨ Á¤º¸¸¦ ±â°è°¡ ÇØµ¶ °¡´ÉÇÑ(machine-readable) RDF Áö½Ä ±×·¡ÇÁ ÇüÅ·Π»óÈ£ ¿¬°áÇÑ´Ù. µû¶ó¼­ ÀÌ Ç¥ÁØÀº AI ¸ðµ¨ Á¤º¸ÀÇ ÆÄÆíÈ­¸¦ ±Øº¹ÇÏ°í »óÈ£¿î¿ë¼ºÀ» Áõ´ë½Ã۸ç, ±Ã±ØÀûÀ¸·Î AI »ýŰèÀÇ Åõ¸í¼º, ½Å·Ú¼º, Àç»ç¿ë¼ºÀ» ³ô¿© ±â¼ú ¹ßÀü¿¡ ±â¿©ÇÏ´Â °ÍÀ» ¸ñÀûÀ¸·Î ÇÑ´Ù.
¿µ¹®³»¿ë¿ä¾à This standard defines a standard ontology schema for the structural description of metadata across the entire lifecycle of various artificial intelligence (AI) models and their related assets, including datasets, documents, and source code. As AI models become increasingly complex and derivative models proliferate, ensuring the provenance, reproducibility, and reliability of each model is essential for the sustainability of the industrial ecosystem. In particular, if core information such as a model's architecture, training and evaluation processes, and utilized datasets is fragmented across different platforms or exists only as unstructured text, objective performance comparisons or searches for models meeting specific criteria become nearly impossible. To address these challenges, this standard proposes a unified metadata structure that encompasses not only the technical specifications of AI models but also the entire process of training, validation, and evaluation, as well as ethical considerations. Through core classes such as AIModel, Dataset, TrainingProcess, and ModelCard, and properties based on standard vocabularies (DCAT, SKOS, Schema.org), information from diverse model sources is interconnected in the form of a machine-readable RDF knowledge graph. Therefore, this standard aims to overcome the fragmentation of AI model information and enhance interoperability, ultimately contributing to the transparency, reliability, and reusability of the AI ecosystem.
±¹Á¦Ç¥ÁØ
°ü·ÃÆÄÀÏ TTAK.KO-10.1594_[1].pdf TTAK.KO-10.1594_[1].pdf            

ÀÌÀü
´ÙÀ½