
ãWIN5ã³ã©ã ãå ãã¬ãŒã¹ãšèããã¬ãŒã¹ãåãAIã«äºæ³ãããŠããïŒ
18çªäººæ°ã®éЬãåã£ãã
ååäžéЬåžã
3é£åã¯ã
ãšãã§ããªãé åœã
ãããªå€§èãã¬ãŒã¹ãèŠããšãã
äžã€çåãæµ®ãã³ãŸãã
ããããã¬ãŒã¹ããAIã«åŠç¿ãããæ¹ãããã®ã§ããããã
ãã€ãéãåŠç¿ãããã°ã
ãã»ãšãã©èµ·ããªãäŸå€ãã«ã
AIãæ¯ãåããããããããªãã
éã«ã
ãããªã®å¶ç¶ã ããé€å€ã
ãšããŠããŸãã°ã
AIã¯ã
倧穎ãåã€äžçãã®ãã®ãã
åŠã°ãªããªããŸãã
ã§ã¯ã
倧èãã¬ãŒã¹ã ãã
å¥ã®AIã«ä»»ããã°ããã®ã§ããããã
äŸãã°ã
å ãã¬ãŒã¹å°éAIã
èããã¬ãŒã¹å°éAIã
ãããŠã
çºèµ°åã«ã
仿¥ã¯ã©ã¡ãã䜿ããæ±ºããã
çŽæçã«ã¯ã
ããªãåççã«èŠããŸãã
ã§ãã
ç«¶éŠ¬ã¯æ¬åœã«ã
ãå ãããšãèãããã®äºçš®é¡
ãªã®ã§ããããã
ä»åã¯ã
倧èãã¬ãŒã¹ãAIã«ã©ãåŠã°ããã¹ããã
ãããŠã
ããããã
äžã€ã®AIã«ãã¹ãŠã®ã¬ãŒã¹ãäºæ³ãããŠããã®ãã
ãèããŸãã
倧èãã¬ãŒã¹ã¯ãééã£ãããŒã¿ãã§ã¯ãªã
ãŸãã
ãããåããå¿ èŠããããŸãã
18çªäººæ°ãåã£ãã
ã ããã
ãã®ããŒã¿ã¯ç°åžžã
ãšããããã§ã¯ãããŸããã
å ¥åãã¹ã§ãã
çµæã®èšé²ãã¹ã§ããªãã
å®éã«ã
ãã®éЬãåã£ãã®ãªãã
ããã¯ã
çããããã©æ£ããããŒã¿
ã§ãã
æ©æ¢°åŠç¿ã§ã¯ã
æ°ã®å°ãªãéèŠãªã±ãŒã¹ãã©ãåŠã¶ããšããã
Class Imbalance
ã®åé¡ãã
é·ãç ç©¶ãããŠããŸãã
倿°æŽŸã®ã±ãŒã¹ãå§åçã«å€ããšã
ã¢ãã«ã¯ã
å°æ°ãããªãã±ãŒã¹ãã
ããŸãåŠã¹ãªãããšããããŸãã
競銬颚ã«ã
ããªãåçŽåããŠèããŠã¿ãŸãã
100ã¬ãŒã¹äžã
1ïœ3çªäººæ°ãåã£ãã¬ãŒã¹ã70ã
4ïœ6çªäººæ°ãåã£ãã¬ãŒã¹ã20ã
ããããäžã®äººæ°éЬãåã£ãã¬ãŒã¹ã10ã
ã ã£ããšããŸãã
æ®éã«åŠç¿ããã°ã
ãäžäœäººæ°ãåã€ã
ãšãããã¿ãŒã³ãã
倧éã«èŠãããšã«ãªããŸãã
ããèªäœã¯ã
ééãã§ã¯ãããŸããã
ã§ãã
æåŸã®10ã¬ãŒã¹ã«ã
äœãå ±éããäºåæ¡ä»¶ããã£ããšããŠãã
70ã¬ãŒã¹åŽã«åãããŠã
ååã«åŠç¿ã§ããªãå¯èœæ§ããããŸãã
ã§ã¯ã倧èãã匷ãåŠç¿ãããã°ããïŒ
ããã§ã
次ã®çºæ³ãåºãŠããŸãã
倧èãã¬ãŒã¹ãå°ãªããªãã
倧èãã¬ãŒã¹ã ã匷ãåŠç¿ãããã°ããã
æ©æ¢°åŠç¿ã«ã¯ã
å°æ°ã±ãŒã¹ãžã
ãã匷ãéã¿ãäžããæ¹æ³ããããŸãã
å°æ°ã±ãŒã¹ãã
åŠç¿æã«å€ãã«èŠããæ¹æ³ããããŸãã
ãŸãã
倧éã®ç°¡åãªã±ãŒã¹ã«åŠç¿ãæ¯é ãããªãããã
é£ããã±ãŒã¹ãžåŠç¿ãéäžãããã
Focal Loss
ã®ãããªæ¹æ³ãç ç©¶ãããŠããŸãã
å ã ã¯ã
ç»åèªèã®ãããªã
競銬ãšã¯å¥åéããçãŸããææ³ã§ãã
ãã ã
èãæ¹ãšããŠã¯ã
倿°æŽŸãæ£è§£ããã ãã§æºè¶³ãããªãã
ãšããçºæ³ã§ãã
ã§ã¯ã
競銬ã§ãã
人æ°èåå©ãã
5åã
10åã
éãåŠç¿ãããã°ããã®ã§ããããã
ããã«ãã
èœãšã穎ããããŸãã
ãç©ŽãæŸããAIããšãæ£ããåçãåºãAIãã¯åãã§ã¯ãªã
äŸãã°ã
å®éã«ã¯ã
人æ°èãåã€å²åãã
5ïŒ ãããªãäžçã ã£ããšããŸãã
ãšãããã
åŠç¿æã«ã
ãã®ã±ãŒã¹ãã
人çºçã«5åã«å¢ãããã
ãããšã
ã¢ãã«ããèŠããäžçã§ã¯ã
人æ°èåå©ãã
çŸå®ããé »ç¹ã«èµ·ããŠããããã«èŠããŸãã
çµæã
æ¬æ¥5ïŒ çšåºŠã§ããã±ãŒã¹ã«ã
15ïŒ ã
20ïŒ ã
ãšã
é«ããã確çãä»ãããããããŸããã
å°æ°ã±ãŒã¹ãå€ãæŸããããã«ã¯ãªã£ãã
ã§ãã
確çãã®ãã®ã¯ã
å£ããŠããã
ãšããããšãããåŸãŸãã
2024幎ã«ã¯ã
äžåè¡¡ããŒã¿ã«å¯Ÿããã
ã©ã³ãã ãªover-samplingãunder-samplingãã
確çäºæž¬ã®Calibrationãžäžãã圱é¿ã調ã¹ãç ç©¶ãçºè¡šãããŠããŸãã
ãã®ç ç©¶ã§ã¯ã
人工çã«ãã©ã³ã¹ãå€ããããŒã¿ã§ã
ãã®ãŸãŸç¢ºçäºæž¬ãè¡ããšã
Calibrationã倧ãã厩ããã±ãŒã¹ã瀺ãããŸããã
ãã¡ããã
ããã¯ç«¶éЬç ç©¶ã§ã¯ãããŸããã
ãã ã
競銬AIã§ãã
åçãã®ãã®ã䜿ããã
ã®ã§ããã°ã
éèŠãªæ³šæã§ãã
çããã±ãŒã¹ãèŠã€ããããããããšã
ãšã
ãã®ã±ãŒã¹ãèµ·ãã確çãæ£ããåºãããšã
ã¯ã
åãã§ã¯ãããŸããã
ããã倧èããæ¶ãã°ããïŒ
ãã¡ããã
åé¡ããããŸãã
倧èãã¬ãŒã¹ãã
ãã¹ãŠé€å€ãããšããŸãã
ãããšã
åŠç¿ããŒã¿ã®äžçããã
倧èããæ¶ããŸãã
çŸå®ã«ã¯ã
ããŸã«äººæ°èãåã€ã
ã§ãã
AIãåŠãã äžçã§ã¯ã
ããããã±ãŒã¹ãååšããªãã
ãã®ç¶æ ã§ã
çŸå®ã®ã¬ãŒã¹ãäºæž¬ããã
ãããšã
äºæž¬ããäžçãšãåŠç¿ããäžçãéãã
ããšã«ãªããŸãã
ããã
倧èããé€å€ãããªãã
ãã®ã¢ãã«ãåŠãã§ããã®ã¯ã
ããã¹ãŠã®ç«¶éЬã
ã§ã¯ãããŸããã
倧èãã±ãŒã¹ãé€ããæ¡ä»¶ä»ãã®ç«¶éЬ
ã§ãã
ã ããã
åçŽã«ã
çããããåé€ã
ã§ã¯æžã¿ãŸããã
ãªããå ãAIããšãèããAIãã«åããïŒ
ããã§ã
å°éã¢ãã«ãšããçºæ³ãåºãŠããŸãã
äŸãã°ã
ã¢ãã«A
å ãã¬ãŒã¹å°éã
äžäœäººæ°ã®äžããã
ã©ã®éЬãäžçªåŒ·ãããã
现ããèŠåããã
ã¢ãã«B
èããã¬ãŒã¹å°éã
人æ°èãŸã§èŠéãåºããŠã
æ®æ®µãšã¯éãæ¡ä»¶ãã
éãèŠãã
ãããŠã
ã¢ãã«C
çºèµ°åã«ã
ãã®ã¬ãŒã¹ãã
å ãåŽãªã®ãã
èããåŽãªã®ãã
ã倿ããã
äžèŠã
ããªãåççã§ãã
æ©æ¢°åŠç¿ã«ã¯ã
ããã«è¿ãã
Mixture of Experts
ãšããèãæ¹ããããŸãã
è€æ°ã®å°éã¢ãã«ã
Expertãçšæããã
ãããŠã
å ¥åã«å¿ããŠã
ã©ã®Expertãžä»»ãããæ±ºããã
1991幎ã«ã¯ãã§ã«ã
è€æ°ã®Expert networkãšã
ãããæ¯ãåããgating networkãçµã¿åãããç ç©¶ãçºè¡šãããŠããŸãã
ã§ããç«¶éŠ¬ã¯æ¬åœã«äºçš®é¡ïŒ
ããã§ã
ç§ã¯å°ãåŒã£ããããŸããã
1çªäººæ°ãåã£ãã
ããã¯ã
å ãåŽã£ãœãã
15çªäººæ°ãåã£ãã
ããã¯ã
èãåŽã£ãœãã
ã§ã¯ã
4çªäººæ°ã¯ïŒ
5çªäººæ°ã¯ïŒ
6çªäººæ°ã¯ïŒ
ããã«ã
åã5çªäººæ°ã§ãã
åå7åã
åå25åã
ã§ã¯ã
åžå Žã§ã®äœçœ®ã¥ãã¯ã
ããªãéããŸãã
8é ç«ãŠã®5çªäººæ°ãšã
18é ç«ãŠã®5çªäººæ°ãã
åããšã¯èšãã«ããã
ã€ãŸãã
人æ°é äœã ãã§ã
å ãã
èããã
ã®äºã€ã«åããšã
ããªãå€ãã®æ å ±ãã
æšãŠãããšã«ãªããŸãã
äžéã®äžçãæ®éã«ãã
äŸãã°ã
ãããªã¬ãŒã¹ãèããŠã¿ãŸãã
1çªäººæ°ã30ïŒ
2çªäººæ°ã25ïŒ
3çªäººæ°ã18ïŒ
4çªäººæ°ã14ïŒ
5çªäººæ°ã8ïŒ
ãã®ä»ã5ïŒ
ãã®ã¬ãŒã¹ã§ã
4çªäººæ°ãåã£ãã
ããã¯ã
倧èãã§ããããã
å°ãªããšãã
18çªäººæ°ãåã£ãã±ãŒã¹ãšã¯ã
ããªãéããŸãã
ã§ãã
ãäžäœ3çªäººæ°ä»¥å€ãåã£ããèãã
ãšå®çŸ©ããã°ã
åãèãåŽã«å ¥ã£ãŠããŸããŸãã
éã«ã
ã7çªäººæ°ä»¥äžã ããèãã
ãªãã
4ïœ6çªäººæ°ã®åå©ã¯ã
å šéšã
å ãåŽã«æŒã蟌ãŸããŸãã
ã©ã¡ããã
å°ãéã§ãã
ãèããããã¯åã¡éŠ¬äººæ°ã ãã§ã決ãŸããªã
ããã«ã
åã¡éЬã ãèŠãŠãã
ã¬ãŒã¹å šäœã®æ§é ã¯åãããŸããã
äŸãã°ã
ã¬ãŒã¹A
1çªäººæ°ã1çã
15çªäººæ°ã2çã
12çªäººæ°ã3çã
ã¬ãŒã¹B
5çªäººæ°ã1çã
1çªäººæ°ã2çã
2çªäººæ°ã3çã
åã¡éŠ¬äººæ°ã ããªãã
ã¬ãŒã¹Bã®æ¹ãã
èããŠããŸãã
ã§ãã
3é£ç³»ã®éЬåžãšããæå³ã§ã¯ã
ã¬ãŒã¹Aã®æ¹ãã
倧ãã厩ããå¯èœæ§ããããŸãã
ã€ãŸãã
ãåã¡éЬãäœçªäººæ°ã ã£ããã
ãšã
ãã¬ãŒã¹å šäœãã©ããããäºæ³å€ã ã£ããã
ãã
åãã§ã¯ãããŸããã
2026幎ã®JRAç ç©¶ã§ããæ³¢ä¹±ããäžã€ã®äººæ°é äœã ãã§å®çŸ©ããŠããªã
2026幎ã«ã¯ã
æ¥æ¬äžå€®ç«¶éЬããŒã¿ã䜿ã£ãŠã
ã¬ãŒã¹åäœã®æ³¢ä¹±ãªã¹ã¯ãã
çºèµ°åæ å ±ããæšå®ããç ç©¶ãçºè¡šãããŠããŸãã
ãã®ç ç©¶ã§ãããæ³¢ä¹±ãã¯ã
åçŽã«ã
6çªäººæ°ä»¥äžãåã£ãã
ã ãã§ã¯ãããŸããã
人æ°èã®åå©ã
ååé«é åœã
人æ°èã®è€ååå ¥ãã
è€åé«é åœã
äžå®ä»¥äžã®é«é¡ææ»ã
ãªã©ãçµã¿åãããã
è€åçãªã©ãã«
ã§ãã
ãããŠã
çºèµ°åã®æ å ±ããã
åã¬ãŒã¹ã«ã€ããŠã
波乱確ç
ãåºããŠããŸãã
ç¬ç«ããæéé ãã¹ã4,556ã¬ãŒã¹ã§ã¯ã
äºæž¬ãããæ³¢ä¹±ç¢ºçãé«ãäžäœ10ïŒ ã§ã
å®éã®è€å波乱äºè±¡ã®çºççã¯ã
çŽ68.2ïŒ ã
äžæ¹ã
äœãåŽ10ïŒ ã§ã¯ã
çŽ21.9ïŒ ã§ããã
éèŠãªã®ã¯ã
0ã1ã ãã§ã¯ãªãã
ãã©ã®ãããæ³¢ä¹±åŽãªã®ãã
ãã
é£ç¶çãªç¢ºçãšããŠæ±ã£ãŠããããšã§ãã
ãå ãïŒèãããã§ã¯ãªããã©ã®ãããèãåŽãã
ããã¯ã
競銬AIãèããããã§ã
ããªãèªç¶ãªçºæ³ã§ãã
äŸãã°ã
ããã¬ãŒã¹ã¯ã
波乱ãªã¹ã¯10ïŒ ã
å¥ã®ã¬ãŒã¹ã¯ã
35ïŒ ã
å¥ã®ã¬ãŒã¹ã¯ã
70ïŒ ã
ãªãã
ç¡çã«ã
å ãïŒ
èããïŒ
ãšã
äºã€ãžåããªããŠããã
4ïœ6çªäººæ°ãåã€ãããªã
äžéã®ã±ãŒã¹ãã
é£ç¶ããäžçã®äžãžçœ®ããŸãã
å°éAIããã©ã¡ããäžã€ãã ãéžã¶å¿ èŠã¯ãªã
Mixture of Expertsã®èãæ¹ã§ãã
å¿ ãã
ã¢ãã«Aã100ïŒ äœ¿çšã
ã¢ãã«Bã¯0ïŒ ã
ãšããå¿ èŠã¯ãããŸããã
äŸãã°ã
説æçšã«ã
3çš®é¡ã®ExpertãçšæãããšããŸãã
Expert A
äžäœéäžåã
Expert B
äžé人æ°ãŸã§åºããåã
Expert C
人æ°èãŸã§å€§ããåºããåã
ããã¬ãŒã¹ã§ã¯ã
A 80ïŒ
B 18ïŒ
C 2ïŒ
å¥ã®ã¬ãŒã¹ã§ã¯ã
A 30ïŒ
B 55ïŒ
C 15ïŒ
ããã«å¥ã§ã¯ã
A 10ïŒ
B 30ïŒ
C 60ïŒ
ãšããããã«ã
è€æ°Expertã®äºæž¬ãã
éã¿ä»ãã§æ··ããããšãã§ããŸãã
ãããªãã
4ïœ6çªäººæ°ãåã€ãããªã
äžéçãªã¬ãŒã¹ãã
ç¡çããã
å ãåŽã倧èãåŽãžã
æŒã蟌ãå¿ èŠã¯ãããŸããã
ãã ããå°éã¢ãã«ãå¢ããã°åŒ·ããªãããã§ããªã
ããã§ã
ãŸãçœ ããããŸãã
å°éã¢ãã«ãäœããšã
æ°ããééããå¢ããŸãã
ãããã
æ¯ãåãééã
ã§ãã
æ¬åœã¯ã
人æ°èãŸã§åºããExpertãã
匷ã䜿ãã¹ãã¬ãŒã¹ã ã£ãã
ã§ãã
gateãã
äžäœéäžExpertã90ïŒ äœ¿ã£ãã
ãããªãã°ã
å°éåããæå³ããããŸããã
ã€ãŸãã
åäžã¢ãã«ãªãã
äºæž¬ã¢ãã«ã®èª€å·®ã ãã ã£ãã
å°éã¢ãã«æ¹åŒã§ã¯ã
ãããžã
ã©ã®ã¢ãã«ã䜿ãããšãã誀差
ãå ãããŸãã
倧èãå°éAIã»ã©ãåŠç¿ããŒã¿ãå°ãªã
ããã«ã
å°éåããã»ã©ã
äžã€ã®ExpertãèŠãããŒã¿ã¯ã
æžããŸãã
ä»®ã«ã
1äžã¬ãŒã¹ãã£ããšããŠãã
æ¬åœã«æ¥µç«¯ãªå€§èããã
200ã¬ãŒã¹ãããªãã£ãã
ãã®200ã¬ãŒã¹ã ãã§ã
è€éãªã¢ãã«ãäœãã°ã
ä»åºŠã¯ã
éåŠç¿
ãèµ·ãããããªããŸãã
éå»ã®200ã¬ãŒã¹ã«ã¯ã
ãã®ãããåãã
ã§ãã
次ã®å€§èãã¬ãŒã¹ã«ã¯ã
ãŸã£ããåããªãã
ãšããããšãããåŸãŸãã
ã ããã
å°éAIãäœãã°ãå°éç¥èãå¢ããŠåŒ·ããªãã
ãšãéããŸããã
ã競銬ã¯3çš®é¡ããšäººéãæ±ºããå¿ èŠããªã
ããã«èšãã°ã
人éãã
æåããã
競銬ã¯3çš®é¡ã§ãã
ãšæ±ºããå¿ èŠããããŸããã
Mixture of Expertsã§ã¯ã
ã©ã®å ¥åé åãã
ã©ã®Expertãæ åœãããèªäœãã
åŠç¿ãããèãæ¹ããããŸãã
ãããšã
çµæãšããŠã
人éããèŠããšã
ãå ãã¬ãŒã¹ã
ãäžéã¬ãŒã¹ã
ãèããã¬ãŒã¹ã
ã«åããããããããŸããã
ã§ãã
ãŸã£ããéãåããæ¹ã«ãªãå¯èœæ§ããããŸãã
äŸãã°ã
å°é æ°ã§èœåå·®ã倧ããã¬ãŒã¹ã
å€é æ°ã§äžäœãæ®æããŠããã¬ãŒã¹ã
ãã³ãæŠã§äžéå±€ãŸã§å¯éããŠããã¬ãŒã¹ã
1é ã ãæããŠããããã©ã
ãã®äžãæ··æŠã®ã¬ãŒã¹ã
å šéŠ¬ãæšªäžŠã³ã®ã¬ãŒã¹ã
AIã«ãšã£ãŠã¯ã
ãã¡ãã®æ¹ãã
èªç¶ãªåãæ¹ãããããŸããã
ãããããèã床ãã¯äžæ¬ã®ç©å·®ãã§æž¬ããã®ã
ãããŸã§æ¥ããšã
ããã«çåãåºãŸãã
äŸãã°ã
2匷å
A 35ïŒ
B 32ïŒ
ãã®ä»33ïŒ
äžéãŸã§æšªäžŠã³å
A 18ïŒ
B 16ïŒ
C 15ïŒ
D 13ïŒ
E 11ïŒ
âŠâŠ
軞å ã»çžææ··æŠå
Aã®åçã¯é«ãã
ã§ãã
2çã»3çåè£ãåºãã
å šäœæ··æŠå
åã¡éЬåè£ãã
2çã»3çåè£ãã
åºãã
å šéšã
ãèã床50ã
ãšäžã€ã®æ°åã«ããŠããŸããšã
åãã«èŠãããããããŸããã
ã§ãã
競銬ãšããŠã¯ã
ããªãéããŸãã
ã€ãŸãã
èããããã¯ãäžã€ã®è»žã ãã§ã¯è¡šããªãå¯èœæ§ãããã
ãšããããšã§ãã
人æ°é äœãããåžå Žå šäœã®åœ¢ããèŠã
ããäžã€ã
èãæ¹ããããŸãã
1çªäººæ°ãã
äœåãã
ã ããèŠãã®ã§ã¯ãªãã
å šéŠ¬ã®åžå Žç¢ºçããã©ã忣ããŠããã
ãèŠãã
äŸãã°ã
åžå Žã®è©äŸ¡ãã
60ïŒ
15ïŒ
10ïŒ
æ®ã15ïŒ
ãªãã
ããªã1é ãžéäžããŠããŸãã
äžæ¹ã
15ïŒ
14ïŒ
13ïŒ
12ïŒ
10ïŒ
âŠâŠ
ãªãã
ããªãåºã忣ããŠããŸãã
ããããéãã¯ã
Entropyãªã©ã
ååžã®åºãããè¡šãææšã§ã
æ°å€åããããšãã§ããŸãã
ã5çªäººæ°ã ããäžç©Žã
ã®ãããªåºå®çãªåŒã³æ¹ããã
åžå Žå šäœãã
äœé ãžã
ã©ãããã忣ããŠããã®ãã
ãèŠãæ¹ãã
é£ç¶çã«æ±ããŸãã
ã§ãåžå Žããçã®èã床ããæããŠãããããã§ããªã
ãã¡ããã
åžå Žãäžèœã§ã¯ãããŸããã
ååãªããºã«ã¯ã
å€ãã®åå è ã®æ å ±ãã
éçŽãããŠããŸãã
ããªã匷ãåºæºã§ãã
ã§ãã
ãªããºã¯ã
çã®åçãã®ãã®ã§ã¯ãããŸããã
競銬åžå Žã§ã¯ã
人æ°éЬãšäººæ°èã®äŸ¡æ Œä»ãã«ã
系統çãªåããååšããã
ããããã
FavoriteâLongshot Bias
ãé·ãç ç©¶ãããŠããŸãã
ã ããã
åžå Žã®éäžåºŠã¯ã
æåãªæ å ±ã
ã§ãã
ããã ãã§ã
ã¬ãŒã¹ã®æ¬åœã®äžç¢ºå®æ§ãã
å®å šã«è¡šããããã§ã¯ãããŸããã
äžçªå±ãªãã®ã¯ãçµæãèŠãŠããäžçãåãããããš
ããã¯ã
絶察ã«å€ããŸããã
çµæãèŠãŠã
18çªäººæ°ãåã£ãã
ã ããã
ãã®ã¬ãŒã¹ã¯å€§èãã¬ãŒã¹ã§ããã
ãããŠã
倧èãå°éAIãªãåœãŠãããŸããã
ãšèšãã
ããã¯ã
äºæž¬ã§ã¯ãããŸããã
åãããªãã
å¿ ãã
çºèµ°åã«åããæ å ±ã ã
ã§ã
ã©ããªã¬ãŒã¹æ§é ãªã®ããã
倿ããå¿ èŠããããŸãã
é æ°ã
ã¯ã©ã¹ã
ãã³ãã
åžå Žã®éäžåºŠã
äžäœè©äŸ¡ã®å·®ã
èè³ªæ§æã
æ¡ä»¶æ¿ããã
ãªã©ã
çµæåã«ååŸå¯èœãªæ å ±ã ãã§ãã
倧èãã«ãªã£ãããã倧èãã¢ãã«ã䜿ãã
ã§ã¯ãªãã
倧èãã«ãªãåã«ã倧èãåŽã®å¯èœæ§ãèªèã§ãããã
ã§ãªããã°ã
æå³ããããŸããã
ã§ã¯ãã©ãæ¯èŒããã°ããïŒ
ããæ¬åœã«ã
å°éã¢ãã«ãæå¹ãªã®ãã調ã¹ããªãã
äžã€ã®ã¢ãã«ã ã詊ããŠã
åœãã£ãã
ã§ã¯è¶³ããŸããã
äŸãã°ã
åãæéé ããŒã¿ã§ã
A
å šã¬ãŒã¹ããã®ãŸãŸåŠç¿ããã
åäžã¢ãã«ã
B
極端ãªã¬ãŒã¹ãé€å€ããã¢ãã«ã
C
çããã±ãŒã¹ãžã
éã¿ãä»ããåäžã¢ãã«ã
D
å ãïŒèãããã
äºååé¡ããŠäœ¿ãåããã¢ãã«ã
E
è€æ°Expertãã
é£ç¶çãªéã¿ã§çµã¿åãããã¢ãã«ã
ãã
åãæªäœ¿çšæéã§æ¯ã¹ãã
ãããŠã
1ççäžçã ãã§ã¯ãªãã
Log Lossã
Brier Scoreã
Calibrationã
人æ°åž¯å¥ã®ææã
é«ãªã¹ã¯åž¯ãšäœãªã¹ã¯åž¯ã§ã®æçžŸã
ãªã©ãèŠãã
穎銬ã ãæŸããããã«ãªã£ãã
ã§ãã
å šäœã®åçããã¡ããã¡ãã«ãªã£ãã
ã§ã¯ã
æ¹åãšã¯èšãã«ããããã§ãã
WIN5ææåã§ããåé¡ããšãå°çšã¢ãã«åãã¯å¥åé¡
WIN5ææåã§ã¯ã
å ¬éäžã
Bæºåºå®ã
èãèã
æ¡åŒµã
hardã
ãªã©ã
ã¬ãŒã¹æ§é ã衚ã蚺æã©ãã«ã䜿ã£ãŠããŸãã
ããã¯ã
ãã¹ãŠã®ã¬ãŒã¹ãã
åãé£ãããšããŠæ±ããªãããã®ãã®ã§ãã
ãã ãã
åé¡ãããããš
ãšã
åé¡ããšã«å¥ã®äºæž¬ã¢ãã«ã䜿ã£ãæ¹ã匷ãããš
ã¯ã
åãã§ã¯ãããŸããã
äŸãã°ã
èãèãšèšºæãããã¬ãŒã¹ã ãã
å¥ã¢ãã«ã«å€ããã°ã
æ¬åœã«è¯ããªãã®ãã
ããã¯ã
å®éã«ã
ååããªæªäœ¿çšããŒã¿ã§æ¯èŒããªããã°åãããŸããã
人éã«ã¯ã
ãå°éã¢ãã«ã®æ¹ãè¯ãããã
ãšèŠããŠãã
åäžã¢ãã«ã®æ¹ãã
å®å®ããŠåŒ·ãå¯èœæ§ããããŸãã
ããã§ãã
æåŸã¯ã
æ€èšŒã§ãã
ãŸãšã
倧èãã¬ãŒã¹ã¯ã
AIã«æããã¹ãã§ããããã
ãŸãã
æ£ãã芳枬ããã倧èãã¬ãŒã¹ãªãã
ããã¯ã
ééã£ãããŒã¿ã§ã¯ãããŸããã
çãããããšãã£ãŠã
åçŽã«åé€ããã°ã
çŸå®ã®ç«¶éЬãšéãäžçãã
AIã«åŠã°ããããšã«ãªããŸãã
ã§ãã
çããã±ãŒã¹ãã
匷ãåŠç¿ãããã°ããã
ãšãéããŸããã
å°æ°ã±ãŒã¹ã人工çã«åŒ·èª¿ããã°ã
ç©ŽéŠ¬ãæŸãããããªãäžæ¹ã§ã
åçãã®ãã®ãã
æªãå¯èœæ§ããããŸãã
ã§ã¯ã
å ãã¬ãŒã¹ãšã
èããã¬ãŒã¹ã§ã
AIãäºã€ã«åããã°ããã®ã§ããããã
ãããã
åçŽã§ã¯ãããŸããã
競銬ã«ã¯ã
4ïœ6çªäººæ°ãåã€ãããªã
äžéã®äžçããããŸãã
ããã«ã
1é ãžäººæ°ãéäžããã¬ãŒã¹ã
äžäœæ°é ãæ®æããã¬ãŒã¹ã
äžé人æ°ãŸã§åºããã¬ãŒã¹ã
軞ã¯å ãããã©ã
çžæãèªããªãã¬ãŒã¹ã
å šäœãæ··æŠã®ã¬ãŒã¹ã
ãªã©ã
ãããããªæ§é ããããŸãã
競銬ã¯ã
ãå ãäžçããšãèããäžçãã®äºçš®é¡ã ãã§ã¯ãªãã
ã®ã ãšæããŸãã
ã ããã
AIã«å¿ èŠãªã®ã¯ã
倧èããšããçµæã ããã
ç¹å¥ã«èŠããããšã§ã¯ãªãã
ç®ã®åã®ã¬ãŒã¹ããã©ããªäžçã«è¿ãã®ãããçµæãåºãåã«èŠåããããšã
ãªã®ãããããŸããã
ãããŠã
å°éã¢ãã«ãäœããšããŠãã
äºæã§åãæ¿ããå¿ èŠã¯ãªãã
è€æ°ã®èŠæ¹ãã
ã¬ãŒã¹ããšã«ã
å°ããã€æ··ãããšããæ¹æ³ããããŸãã
æçµçã«éèŠãªã®ã¯ã
倧èããåŠç¿ããããã
ã§ã¯ãªãã
倧èããèµ·ãããããæ¡ä»¶ããçºèµ°åã«æãããããã
ã§ãã
ãäŸå€ããèŠããããšããã
äŸå€ãèµ·ãããããäžçãèŠã€ããããšã
競銬AIã«ãšã£ãŠã
æ¬åœã«å¿ èŠãªã®ã¯ã
ãã¡ããªã®ãããããŸããã
泚æäºé
æ¬èšäºã§ã¯ããå ãã¬ãŒã¹ããèããã¬ãŒã¹ããäžé人æ°åž¯ããªã©ã®è¡šçŸã説æäžäœ¿çšããŠããŸããããããã«ç«¶éŠ¬å šäœã§çµ±äžãããå³å¯ãªå®çŸ©ããããšããæå³ã§ã¯ãããŸããã
4ïœ6çªäººæ°ãªã©ãäžåŸã«ã穎銬ããäžç©Žããšå®çŸ©ããŠããããã§ããããŸããã人æ°é äœã¯é æ°ããªããºååžã«ãã£ãŠæå³ãå€ãããããæ¬èšäºã§ã¯åºå®çãªåºåãšããŠæ±ã£ãŠããŸããã
Class ImbalanceãFocal LossãMixture of Expertsãªã©ã¯äžè¬çãªæ©æ¢°åŠç¿ç ç©¶ã®èãæ¹ã競銬ãžåœãŠã¯ããŠèª¬æãããã®ã§ãããããã®ææ³ã競銬ãžå°å ¥ããã°ãäºæž¬æ§èœãæ¹åãããšå®èšŒããããšããæå³ã§ã¯ãããŸããã
å°æ°ã±ãŒã¹ãžã®éã¿ä»ããresamplingã¯ãå°æ°äºè±¡ã®æ€åºæ§èœãæ¹åããå ŽåããããŸããã確çäºæž¬ã®CalibrationãæãªãããšããããŸããåçãå©çšããã¢ãã«ã§ã¯ãè奿§èœã ãã§ãªãCalibrationãå¥éè©äŸ¡ããå¿ èŠããããŸãã
2026幎ã®JRAããŒã¿ã䜿ã£ã波乱ãªã¹ã¯ç ç©¶ã§çšãããããæ³¢ä¹±ãã¯ã人æ°èã®åå©ã ãã§ã¯ãªãã人æ°èã®è€ååå ¥ããé«é åœãªã©ãçµã¿åãããç ç©¶ç¬èªã®è€åã©ãã«ã§ããäžäœ10ïŒ ã§èŠ³æž¬ããã68.2ïŒ ãšããæ°å€ã¯ãã人æ°èã®éЬã68.2ïŒ ã®ç¢ºçã§åã€ããšããæå³ã§ã¯ãããŸããã
ãŸããçµæåŸã«ã¬ãŒã¹ããå ãããèããããšåé¡ããŠãããã«é©ããã¢ãã«ãéžã¶ããšã¯ãäºåäºæž¬ã®æ€èšŒã«ã¯ãªããŸãããå°éã¢ãã«ã䜿ãåããå Žåããã¢ãã«éžæã«äœ¿ãæ å ±ã¯çµæåã«å©çšå¯èœãªãã®ãžéå®ããå¿ èŠããããŸãã
äž»ãªåèè³æ
Haibo He & Edwardo A. GarciaïŒ2009ïŒ
âLearning from Imbalanced Dataâ
å°æ°ã¯ã©ã¹ã倿°ã¯ã©ã¹ã«åãããClass Imbalanceåé¡ã«ã€ããŠã代衚çãªææ³ãè©äŸ¡æ¹æ³ãæŽçããã¬ãã¥ãŒã
https://doi.org/10.1109/TKDE.2008.239
Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He & Piotr DollárïŒ2017ïŒ
âFocal Loss for Dense Object Detectionâ
倧éã®ç°¡åãªå€æ°æŽŸãµã³ãã«ãåŠç¿ãæ¯é ããåé¡ã«å¯Ÿããé£ããäŸãžåŠç¿ãéäžãããFocal Lossãææ¡ããç ç©¶ã
https://openaccess.thecvf.com/content_iccv_2017/html/Lin_Focal_Loss_for_ICCV_2017_paper.html
Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan & Geoffrey E. HintonïŒ1991ïŒ
âAdaptive Mixtures of Local Expertsâ
è€æ°ã®Expert networkãšgating networkãçµã¿åãããå ¥åã«å¿ããŠç°ãªãå°éã¢ãã«ãæ åœãããMixture of Expertsã®ä»£è¡šçåæç ç©¶ã
https://doi.org/10.1162/neco.1991.3.1.79
Marco Piccininni et al.ïŒ2024ïŒ
âUnderstanding random resampling techniques for class imbalance correction and their consequences on calibration and discrimination of clinical risk prediction modelsâ
Class Imbalanceã«å¯Ÿããrandom over-samplingïŒunder-samplingãã確çäºæž¬ã®Calibrationãžäžãã圱é¿ãæ€èšããç ç©¶ã競銬ç ç©¶ã§ã¯ãªããããæ¬èšäºã§ã¯äžè¬çãªç¢ºçäºæž¬äžã®æ³šæãšããŠåç §ã
https://doi.org/10.1016/j.jbi.2024.104666
Shuichi SugiuraïŒ2026ïŒ
âLeakage-aware race-level upset-risk diagnostics for prospective horse-race predictionâ
æ¥æ¬äžå€®ç«¶éЬããŒã¿ã䜿ããçµæåã«å©çšå¯èœãªã¬ãŒã¹æ§é æ å ±ãããã¬ãŒã¹åäœã®è€åçãªæ³¢ä¹±ãªã¹ã¯ãæšå®ããç ç©¶ãç¬ç«æéé ãã¹ã4,556ã¬ãŒã¹ã§ãäºæž¬æ³¢ä¹±ãªã¹ã¯äžäœ10ïŒ ãšäžäœ10ïŒ ã®å®æž¬çºççã«å·®ã確èªãããŠããã
https://doi.org/10.3389/frai.2026.1922250
#WIN5 #競銬 #ç«¶éŠ¬äºæ³ #競銬AI #AI競銬 #æ©æ¢°åŠç¿ #倧èã #穎銬 #MixtureOfExperts #ããŒã¿åæ