Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
º£ÀÌÁö¾È ½Å°æ¸ÁÀ» ÀÌ¿ëÇÑ ¼øÂ÷Àû ´ÙÁßÀÛ¾÷ ÇнÀ ÃÖÀûÈ |
¿µ¹®Á¦¸ñ(English Title) |
Sequential Multitask Learning Optimization Using Bayesian Neural Network |
ÀúÀÚ(Author) |
¼Õ¼ºÈ£
±èÁö¼·
À庴Ź
Seongho Son
Jiseob Kim
Byoung-Tak Zhang
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 24 NO. 05 PP. 0251 ~ 0255 (2018. 05) |
Çѱ۳»¿ë (Korean Abstract) |
ÀΰøÁö´ÉÀÇ È°¿ë¼ºÀ» ³ôÀ̱â À§Çؼ´Â ´Ù¾çÇÑ Á¾·ùÀÇ ÀÛ¾÷À» ¼øÂ÷ÀûÀ¸·Î ÇнÀÇÒ ¼ö ÀÖ´Â ´É·ÂÀ» °³¹ßÇÏ´Â °ÍÀÌ ÇʼöÀûÀÌ´Ù. ±×·¯³ª ±âÁ¸ÀÇ ½Å°æ¸ÁÀº ¼øÂ÷Àû ´ÙÁßÀÛ¾÷ ÇнÀ ȯ°æ¿¡¼ ÀÌÀü¿¡ ÇнÀÇß´ø ÀÛ¾÷ÀÇ ¼º´ÉÀÌ ±Þ°ÝÇÏ°Ô ¶³¾îÁö´Â ´ÜÁ¡À» °¡Áö°í ÀÖ´Ù. º£ÀÌÁö¾È ½Å°æ¸ÁÀº °¢ ¿¬°á °¡ÁßÄ¡¸¦ È®·üºÐÆ÷¿¡ ´ëÀÀ½ÃÄÑ Ã¼°èÀûÀÎ ¿Â¶óÀÎ ÇнÀÀÌ °¡´ÉÇÑ ÇÑÆí ¸ðµ¨ÀÇ µ¥ÀÌÅÍ¿¡ ´ëÇÑ ºÒÈ®½Ç¼ºÀ» °è·®ÇÒ ¼ö ÀÖ´Ù. º» ¿¬±¸¿¡¼´Â º£ÀÌÁö¾È ½Å°æ¸ÁÀ» ÀÌ¿ëÇÏ¿© ¼øÂ÷Àû ´ÙÁßÀÛ¾÷ ÇнÀ ȯ°æ¿¡¼ »õ·Î¿î ÀÛ¾÷¿¡ ´ëÇÑ ÇнÀÀ» ¿øÈ°ÇÏ°Ô ¼öÇàÇÔ°ú µ¿½Ã¿¡ ÀÌÀü¿¡ ÇнÀÇß´ø ÀÛ¾÷¿¡ ´ëÇÑ ¼º´ÉÀ» Àß À¯ÁöÇÏ´Â ¸ðµ¨À» Á¦½ÃÇÑ´Ù. Á¦½ÃµÈ ¸ðµ¨Àº º£ÀÌÁö¾È ¿ªÀüÆÄ ±â¹ýÀ» ÀÌ¿ëÇØ °è»êÇÑ °¡ÁßÄ¡ ºÐÆ÷ Æò±ÕÀÇ °æ»çµµ¿¡ »çÀüºÐÆ÷ÀÇ Ç¥ÁØÆíÂ÷¸¦ °öÇÑ °ÍÀ» ÇнÀ¿ë °æ»çµµ·Î ÀÌ¿ëÇÑ´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
It is essential to develop the ability to sequentially learn various tasks, in order to increase the availability of artificial intelligence applications. However, typical neural network models show a drastic decline of the previously learnt task¡¯s performance in a sequential multitask learning environment. Bayesian neural networks can measure the uncertainty of a model on the given data, while enabling systematic online learning by matching each weight to a probability distribution. This research proposes a model using a Bayesian neural network, which maintains its performance on previously learnt tasks while continuously learning new tasks in a sequential multitask learning setting. The proposed model uses Bayesian backpropagation, where the gradient of each weight distribution¡¯s mean is multiplied by the standard deviation of the corresponding prior distribution.
|
Å°¿öµå(Keyword) |
µö·¯´×
°¡ÁßÄ¡ ºÒÈ®½Ç¼º
º£ÀÌÁö¾È ½Å°æ¸Á
¼øÂ÷Àû ´ÙÁßÀÛ¾÷ ÇнÀ
deep learning
weight uncertainty
Bayesian neural network
sequential multitask learning
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|