Ȩ > Ç¥ÁØÈ °³¿ä > TTAÀÇ Ç¥ÁØÇöȲ
Ç¥ÁعøÈ£ | TTAK.KO-10.1558 | ±¸Ç¥ÁعøÈ£ | |
---|---|---|---|
Á¦°³Á¤ÀÏ | 2024-12-06 | ÃÑÆäÀÌÁö | 36 |
ÇѱÛÇ¥Áظí | »ý¼ºÇü ÀΰøÁö´É ÇнÀÀ» À§ÇÑ ¸ÖƼ¸ð´Þ µ¥ÀÌÅÍÀÇ Ç°Áú°ËÁõ ¹æ¹ý | ||
¿µ¹®Ç¥Áظí | The Method for Quality Verification of Multi-Modal Data for Generative AI Learning | ||
Çѱ۳»¿ë¿ä¾à | ÀÌ Ç¥ÁØÀº ¸ÖƼ¸ð´Þ µ¥ÀÌÅÍÀÇ Ç°ÁúƯ¼ºÀΠǥÇö¼º, º¯È¯(ÀÏÄ¡)¼º, Á¤·Ä(°ü°è)¼ºÀ» Á¦½ÃÇϰí, ÀÌ·¯ÇÑ Ç°ÁúƯ¼º¿¡ ±â¹ÝÇÑ ¸ÖƼ¸ð´Þ µ¥ÀÌÅÍÀÇ ´ëÇ¥ ÇнÀ¾÷¹«º° ÀǹÌÁ¤È®¼º °ËÁõ Ç׸ñ µµÃâ¹æ¹ýÀ» Á¦°øÇÑ´Ù. ´ëÇ¥ ÇнÀ¾÷¹«´Â ½Ã°¢Àû À½¼º ÀνÄ, ½Ã°¢Àû À½¼º ÇÕ¼º, °´Ã¼/Çൿ ºÐ·ù ¹× ŽÁö, À̺¥Æ® ŽÁö, ÀÌ»ó ŽÁö, °¨Á¤ ÀνÄ, À̹ÌÁö ¼³¸í, ºñµð¿À ¼³¸í, ½Ã°¢Àû ÁúÀÇÀÀ´ä, ¿µ»ó ¿ä¾à, ¿µ»ó °Ë»ö, ¿µ»ó »ý¼º µîÀÌ´Ù. ¶ÇÇÑ ¸ÖƼ¸ð´Þ µ¥ÀÌÅ͸¦ ÇнÀÇÏ¿© Àΰ£ÀÇ º¹ÇÕÁö´É ÆÇ´Ü ´É·Â°ú °°Àº ±â´ÉÀ» ¼öÇàÇÒ ¼ö ÀÖ´Â »ý¼ºÇü ÀΰøÁö´É ¸ÖƼ¸ð´Þ ¸ðµ¨ÀÇ ¼º´ÉÀ» Æò°¡Çϱâ À§ÇÑ À¯È¿¼º ÁöÇ¥ ¹× ÃøÁ¤ »ê½Ä µî ¸ÖƼ¸ð´Þ µ¥ÀÌÅÍÀÇ Ç°ÁúÀ» Æò°¡ÇÏ´Â µ¥ ÇÊ¿äÇÑ °ËÁõ ¹æ¹ýÀ» Á¦°øÇÑ´Ù. | ||
¿µ¹®³»¿ë¿ä¾à | We present the quality characteristics of multimodal data such as Representation, Translation(matching), and Alignment(relationship), and provide a method for deriving semantic accuracy verification items for each representative learning task of multimodal data based on these quality characteristics. There are a total of 12 representative learning tasks, including visual speech recognition, visual speech synthesis, class classification, event detection, anomaly detection, emotion recognition, image description, video description, visual question and answering, image summary, image search, and image generation. In addition, it should be able to perform functions similar to human complex intelligence judgment ability by learning multimodal data. Thus it provides verification methods necessary to evaluate the quality of multimodal data, including validity indicators and measurement formulas to evaluate the performance of generative artificial intelligence multimodal models. | ||
±¹Á¦Ç¥ÁØ | |||
°ü·ÃÆÄÀÏ |
![]() |