Ȩ > Ç¥ÁØÈ °³¿ä > TTAÀÇ Ç¥ÁØÇöȲ
Ç¥ÁعøÈ£ | TTAK.KO-11.0280-Part1 | ±¸Ç¥ÁعøÈ£ | |
---|---|---|---|
Á¦°³Á¤ÀÏ | 2020-12-10 | ÃÑÆäÀÌÁö | 21 |
ÇѱÛÇ¥Áظí | °ËÁõ¿ë µ¥ÀÌÅͼ¼Æ®ÀÇ ¹ë·±½º ±â¹Ý ÀΰøÁö´É ¼ÒÇÁÆ®¿þ¾î ½Å·Ú¼º Æò°¡ ¹æ¹ý - Á¦1ºÎ: ¹æ¹ý·Ð ¹× ü°è | ||
¿µ¹®Ç¥Áظí | A Method for Evaluating the Reliability of Artificial Intelligence Software Based on the Balance of the Validation Dataset : Part 1 - Methodology and System | ||
Çѱ۳»¿ë¿ä¾à | ÃÖ±Ù »ê¾÷º° ´Ù¾çÇÑ ¼ºñ½º¿¡ ÀΰøÁö´ÉÀÌ Æ÷ÇÔµÈ ¼ÒÇÁÆ®¿þ¾î µµÀÔÀÌ È®»êµÇ°í ÀÖÀ¸¸ç, ÀÌ·¯ÇÑ ¼ºñ½ºÀÇ Ç°ÁúÀ» Á¿ìÇÏ´Â ÇÙ½É ¿äÀÎÀº ¼ÒÇÁÆ®¿þ¾î¿¡ žÀçµÈ ÀΰøÁö´ÉÀÇ Á¤È®¼ºÀÌ´Ù. Á¤È®¼ºÀ» °ËÁõÇÏ´Â ÀϹÝÀûÀÎ ¹æ¹ýÀ¸·Î, ÇнÀ¿ë µ¥ÀÌÅͼ¼Æ® ¿Ü, º°µµÀÇ °ËÁõ¿ë µ¥ÀÌÅͼ¼Æ®¸¦ ±¸¼ºÇÏ¿©, ÀΰøÁö´É µ¿ÀÛ °á°úÀÇ Ãâ·Â °ª°ú ºñ±³ÇÏ´Â ¹æ½ÄÀ» »ç¿ëÇÑ´Ù. ÀÌ ¶§ °ËÁõ¿ë µ¥ÀÌÅÍÀÇ Ç¥º» °ÝÂ÷·Î ÀÎÇÏ¿© ƯÁ¤ ¿µ¿ªÀÇ µ¥ÀÌÅÍ°¡ Áö³ªÄ¡°Ô ÆíÁßµÇ¾î »ý»êµÇ°Å³ª, ȤÀº µ¥ÀÌÅÍ ¼öÁý °úÁ¤¿¡¼ÀÇ ÇнÀ µ¥ÀÌÅÍ°¡ ¹èÆ÷µÈ ÀΰøÁö´É ¼ÒÇÁÆ®¿þ¾îÀÇ ¿î¿µ µ¥ÀÌÅ͸¦ Á¦´ë·Î ´ëÇ¥ÇÏÁö ¸øÇÏ°í ƯÁ¤ ¿µ¿ªÀÌ °ú´ë ´ëÇ¥µÇ´Â °æ¿ì°¡ ¹ß»ýÇϸé, ÀÌ°ÍÀ» µ¥ÀÌÅͼ¼Æ®ÀÇ ¹ë·±½º°¡ ÀûÀýÇÏÁö ¸øÇÏ¿© ¹ß»ýµÈ »ùÇøµ ÆíÇâ(sampling bias), »ùÇøµ ¿À·ù(sampling error)¶ó°í ÇÑ´Ù. ÀÌ »ùÇøµ ÆíÇâÀº ÀΰøÁö´É ¼ÒÇÁÆ®¿þ¾îÀÇ Á¤È®¼º °ËÁõ °á°ú¸¦ ¿Ö°î½ÃÅ°´Â ÁÖ¿ä ¿øÀÎÀ̱⠶§¹®¿¡, Çö½Ç ¼¼°èÀÇ ´Ù¾çÇÑ ½Ã³ª¸®¿À¸¦ ó¸®ÇÒ ¼ö ÀÖ´Â Áö¸¦ Æò°¡ÇÏ´Â °üÁ¡¿¡¼´Â, ¹ë·±½º°¡ È®º¸µÈ Æò°¡¿ë µ¥ÀÌÅͼ¼Æ®¸¦ È°¿ëÇÏ¿© ½Å·Ú ¼öÁØÀ» Æò°¡ÇÏ´Â °ÍÀÌ ÇÊ¿äÇÏ´Ù.
ÀÌ¿¡ º» Ç¥ÁØ¿¡¼´Â Æò°¡¿ë µ¥ÀÌÅͼ¼Æ®¸¦ ±¸ÃàÇÏ´Â °úÁ¤¿¡¼, ÀÔ·Â µ¥ÀÌÅÍÀÇ Á¶°Ç¿¡ ºÎÇյǴ µ¥ÀÌÅͼ¼Æ®ÀÇ ºÐÆ÷¸¦ °í¸£°Ô Æí¼ºÇÏ°í, ÀΰøÁö´É ¸ðµ¨ÀÇ ¼ÒÇÁÆ®¿þ¾îÀÇÀ» Æò°¡ÇÏ´Â ¹æ¹ý ¹× ÀýÂ÷¸¦ Á¤ÀÇÇϸç, ¡®ÀΰøÁö´É ¼ÒÇÁÆ®¿þ¾îÀÇ ½Å·Úµµ(AI Reliability)¡¯¶ó´Â ÁöÇ¥¸¦ Á¤ÀÇÇÑ´Ù. |
||
¿µ¹®³»¿ë¿ä¾à | Recently, the introduction of software including artificial intelligence in various services by industry is spreading, and the key factor that determines the quality of these services is the accuracy of artificial intelligence embedded in the software. As a general method of verifying accuracy, a separate verification dataset is constructed in addition to the training dataset and compared with the output value of the artificial intelligence action result. At this time, when the data in a specific area is produced due to the sample gap of the verification data, or when the training data in the data collection process is not properly represented and the operation data of the distributed artificial intelligence software is not properly represented and the specific area is over-represented. When occurs, this is called a sampling bias or sampling error caused by an improper balance of the dataset. Since this sampling bias is a major cause of distorting the accuracy verification results of artificial intelligence software, from the viewpoint of evaluating whether it can handle various real-world scenarios, the confidence level is evaluated using a balanced evaluation dataset. It is necessary to do.
Therefore, in the process of constructing the evaluation dataset, this standard evenly organizes the distribution of the dataset that meets the conditions of the input data, and defines the method and procedure for evaluating the software significance of the artificial intelligence model. It defines an index called 'AI Reliability'. |
||
±¹Á¦Ç¥ÁØ | |||
°ü·ÃÆÄÀÏ | TTAK.KO-11.0280-Part1.pdf |