Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
FedPress: ¿¬ÇÕÇнÀ ±â¹Ý ¹Ì¼¼Á¶Á¤À» ÅëÇÑ È¿À²ÀûÀÌ°í °³ÀÎÁ¤º¸¸¦ °øÀ¯ÇÏÁö ¾Ê´Â ½Å°æ¸Á ¾ÐÃà±â¹ý |
¿µ¹®Á¦¸ñ(English Title) |
FedPress: Efficient Model Compression without Sharing Private Data via Federated Fine-Tuning |
ÀúÀÚ(Author) |
ÀÌ°ÇÈ£
±è¼º¿õ
±èÇöÁØ
±è¹Î¼ö
Á¤Çý¹Î
±è´öȯ
ÃÖµ¿¿Ï
KeonHo Lee
SeongWoong Kim
HyunJun Kim
MinSoo Kim
Heymin Jeong
DeokHwan Kim
DongWan Choi
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 48 NO. 02 PP. 0617 ~ 0619 (2021. 12) |
Çѱ۳»¿ë (Korean Abstract) |
°¡Àå ´ëÇ¥ÀûÀÎ ½Å°æ¸Á °æ·®È ±â¹ý Áß ÇϳªÀÎ ³×Æ®¿öÅ© ÇÁ·ç´×(network pruning)Àº ¾ÐÃà·üÀÌ ³ô¾ÆÁü¿¡ µû¶ó Ã߷п¡ ÇʼöÀûÀÎ ÆĶó¹ÌÅÍ°¡ Á¦°ÅµÇ¾î ¼º´É Ç϶ôÀÌ ¹ß»ýÇÑ´Ù. ÀÌ·¸°Ô ÇÁ·ç´× ÀÌÈÄ Ç϶ôµÈ ¼º´ÉÀº ¹Ì¼¼Á¶Á¤(finetuning)À» ÅëÇØ º¹±¸µÈ´Ù. ÇÏÁö¸¸, Ãֱ٠Ŭ¶óÀ̾ðÆ®ÀÇ °³ÀÎÁ¤º¸ º¸È£°¡ Áß¿äÇØÁö°í µ¥ÀÌÅÍ ¼¹ö À¯Áö ºñ¿ëÀÌ Áõ°¡ÇÔ¿¡ µû¶ó µ¥ÀÌÅ͸¦ ÀúÀåÇÏ°í ¼öÁýÇÏ´Â µ¥ Á¦¾àÀÌ µû¸£±â ¶§¹®¿¡ ¹Ì¼¼Á¶Á¤À» ¼öÇàÇϱ⠾î·Á¿î °æ¿ì°¡ ¸¹´Ù. ÀÌ¿¡ º» ³í¹®¿¡¼´Â ÇÁ·ç´×À¸·Î ¾ÐÃàµÈ ÇÕ¼º°ö ½Å°æ¸ÁÀÇ Ç϶ôµÈ ¼º´ÉÀ» ¼¹ö¿¡¼ÀÇ ¹Ì¼¼Á¶Á¤ °úÁ¤ ¾øÀÌ º¹±¸ÇÏ´Â ¹æ¹ýÀ» Á¦¾ÈÇÑ´Ù. ÀÏÂ÷ÀûÀ¸·Î µ¥ÀÌÅÍ¿Í Ãß°¡ÇнÀ ¾øÀÌ ¼¹ö¿¡¼ Ãʱ⠾ÐÃàµÈ ¸ðµ¨À» º¹±¸ÇÏ°í, ÀÌÈÄ ¼¹ö·ÎÀÇ µ¥ÀÌÅÍ Àü¼Û ¾øÀÌ ¿©·¯ Ŭ¶óÀ̾ðÆ®µéÀÇ ½Å°æ¸ÁÀ» °øÀ¯ÇÏ´Â ¿¬ÇÕ ÇнÀ(Federated Learning)À» ¼öÇàÇÔ À¸·Î½á ¼º´ÉÀ» ´õ¿í Çâ»ó½ÃÅ°´Â ¹æ¹ýÀ» Á¦¾ÈÇÑ´Ù. ResNet-34¸¦ ÀÌ¿ëÇÑ À̹ÌÁö³Ý ½ÇÇè¿¡¼ ÀÏÂ÷ÀûÀÎ ¼º´É º¹±¸¸¦ ¼öÇàÇÑ °æ·® ½Å°æ¸ÁÀÇ ¼º´ÉÀÌ ±âÁ¸ ÇÁ·ç´×¸¸ ¼öÇàÇÑ ½Å°æ¸ÁÀÇ Á¤È®µµ 17%º¸´Ù 30% ´õ ³ôÀº 47% Á¤È®µµ¸¦ ´Þ¼ºÇÏ¿´À¸¸ç, ÀÌÈÄ ¿¬ÇÕ ÇнÀÀ» ÅëÇØ 17.1% ´õ ³ôÀº 64.5% Á¤È®µµ±îÁö ´Þ¼ºÇÔÀ» È®ÀÎÇÏ¿´´Ù. À̶§ °æ·® ½Å°æ¸ÁÀÇ »çÀü ÇнÀµÈ ½Å°æ¸ÁÀÇ FLOPs°ú ÆĶó¹ÌÅÍÀÇ ¼ö´Â »çÀü ÇнÀµÈ ½Å°æ¸Á ´ëºñ ¾à 29% ´õ Àû´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
|
Å°¿öµå(Keyword) |
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|