æèãã°ã©ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ããéçãã¦ç¬¬ 5 å·ã¨ãªãã¾ãããçãã¾ãããã¨ããããã¾ãï¼
ãéçéå ±ððððã
— è¬è«ç¤¾ãµã¤ã¨ã³ãã£ãã£ã¯ðï¸ð (@kspub_kodansha) 2024å¹´8æ1æ¥
æ©æ¢°å¦ç¿ãããã§ãã·ã§ãã«ã·ãªã¼ãºã®éçã決ã¾ãã¾ããðãæèªãããã¨ããããã¾ãâ¼ï¸
æ¾äºå太ã»çè°·äºã転移å¦ç¿ãã4å·ã https://t.co/Qic24KAwxD
ä½è¤ç«é¦¬ãã°ã©ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ãã5å·ã https://t.co/Peqn1ZQavo pic.twitter.com/VBkNp2Uwjj
æ¡æ£ã¢ãã«ã¨æé©è¼¸éï¼æé©è¼¸é第 5 å·ï¼ã GNN ã®ææ°ååï¼ã°ã©ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ç¬¬ 3 å·ï¼ã§ãããã¾ããããéçã«æè¬ãã¦æ¸ãä¸ããå°éè¨äºãæ稿ãã¾ãã
æ¬ç¨¿ã§ã¯æ·±å±¤å¦ç¿ã§é¨å空éãæ±ãã¨ãã¯å°å½±è¡åãèããã¨ããã¨ãã話ããã¾ãã
ç´¹ä»ããåºæ¬çãªã¢ã¤ãã¢ã¯ On the Expressive Power of Spectral Invariant Graph Neural Networks [Zhang+ ICML 2024] ã«åºã¥ãã¦ãã¾ããåç´ãªè©±ãªã®ã§ããã奥深ããå ¼ãæããè¯ã話ãªã®ã§åãä¸ãããã¨ã«ãã¾ããã
ç®æ¬¡
- ç®æ¬¡
- é¨å空éãèããã¢ããã¼ã·ã§ã³
- é¨å空éãæ±ãé£ãã
- é¨å空éã¯å°å½±è¡åã§è¡¨ç¾ããã¨ãã
- ãããã«
é¨å空éãèããã¢ããã¼ã·ã§ã³
ãã¼ã¿ãé¨å空éã¨ãã¦è¡¨ããã¨ããã°ãã°ããã¾ãã
ã¨ã¦ãåãããããä¾ã¯åèªãã¯ãã«ã®éåã§ã [Ishibashi+ NAACL 2024]ããããããã¨ã§ãåéåãå ±éé¨åãªã©ã®æä½ãèªç¶ã«è¡ããã»ããéåã©ããã®é¡ä¼¼åº¦ãå¹³é¢ã¨ãã¯ãã«ã®ã³ãµã¤ã³é¡ä¼¼åº¦ãç¨ãã¦ãããã«å®ç¾©ãããã¨ãã§ãã¾ãã
ä¾ãã°ã"We are the king and queen" 㨠"We are the royalty" ã¨ããæã¯ä¼¼ã¦ãã¾ãããåèªãã¯ãã«ã®ãªã¹ãã¨ãã¦åç´ã«æ¯è¼ããã¨é ãã¨å¤å®ãã¦ãã¾ãã¾ããé¨å空éã¨ãã¦è¡¨ç¾ãããã¨ã§ã"king and queen" ã表ãå¹³é¢ã¨ "royalty" ã®æ¹åã沿ãã®ã§ãè¯ãæãã«æ¯è¼ã§ããããã«ãªãã¾ãã
å¥ã®ä¾ã¯ã°ã©ãä¸ã®ä¿¡å·ã表ãé¨å空éã§ããã°ã©ãã¯ãã°ã©ãä¸ã®æ»ãããªä¿¡å·ã表ãé¨å空éãããã°ã©ãä¸ã®ãã¤ã¸ã¼ãªä¿¡å·ã表ãé¨å空éããªã©ãç¨ãã¦è¡¨ç¾ã§ãã¾ãããã°ã©ãä¸ã®æ»ãããªä¿¡å·ã表ãé¨å空éãã¯ã°ã©ãã®å ¨ä½åããã¾ã表ç¾ã§ãã¦ããã®ã§ã辺ã®éåãªã©å±æçãªè¦ç´ ã®éåã§è¡¨ããããæ§è½ããããªããã¨ããã°ãã°ããã¾ãã
ä¾ãã°ãã°ã©ããå ¥åã¨ãã¦åãåã£ãã¨ãããã®ã°ã©ãä¸ã§ã®æ»ãããªä¿¡å·ã®é¨å空éãè¨ç®ãã¦ããã®é¨å空éãã°ã©ãã®ç¹å¾´ã¨ãããããå¾æ®µã®ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã«æ¸¡ããã¨ãã£ãå¦çãèãããã¾ãããããããã¨ã§ãå¾æ®µã®ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã¯å ¥åã°ã©ãã«ã©ãããã¯ã©ã¹ã¿æ§é ãããããªã©ãå©ç¨ãã¦ã¿ã¹ã¯ã解ããããã«ãªãã¾ãã
é¨å空éãæ±ãé£ãã
ãããã深層å¦ç¿ã§é¨å空éãæ±ãã®ã¯ä¸çç¸ã§ã¯ããã¾ããã大ããªåé¡ã¯ãé¨å空éãè¨ç®æ©ä¸ã§è¡¨ç¾ããã®ãé£ããã¨ãããã¨ã§ãã
1 次å é¨å空éã®å ´åãã ã®æ¹åãã¨ããããã«ã代表çãªåä½ãã¯ãã« ãç¨ãã¦è¡¨ç¾ãããã¨ãèãããã¾ãããããããã®è¡¨ç¾ã¯ä¸æã§ã¯ããã¾ããããã® 1 次å 空éã¯ã ã®æ¹åãã¨ãã¦ã表ç¾ã§ãã¾ãã
ä¾ãã°ã3 次å 空éã®ãã¡ã® 1 次å é¨å空éãèããã¨ããã ã®æ¹åãã¨ããé¨å空éã¯ãã ã®æ¹åãã¨ã表ãããã¨ãããã¨ã§ãã
ãã®è¡¨ç¾ãå¾æ®µã®ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã«æ¸¡ããã¨ãã¾ãããããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã¯ãã¯ãã« ãåãåã£ãã¨ãã¨ããã¯ãã« ãåãåã£ãã¨ãã§ãç°ãªãå¿çãè¿ãããããã¾ããããã® 1 次å é¨å空éã®å¿çã¨ãã¦ã¯ãã©ã¡ãã使ãã°ããã®ã§ãããããã¾ããããã¾ã§ã¯å ¥åã¨ãã¦é¨å空éãä¸ããå ´åãèãã¦ãã¾ããããé¨å空éãäºæ¸¬ãããï¼åºåãããï¼ã¨ããã¨ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã¸ã®æ師ãã¼ã¿ã¨ãã¦ã¯ 㨠ã®ã©ã¡ãã使ãã°ããã§ãããããä½ãèããã«ä¸¡æ¹ãè¨ç·´ãã¼ã¿ã«å ¥ãã¦äºä¹èª¤å·®ã§è¨ç·´ãã¦ãã¾ãã¨ããã®å¹³åã§ããã¼ããã¯ãã«ãåºåããããã«ãªã£ã¦ãã¾ãã¾ãã
ããã¯ã1 次å é¨å空éã®æ¹åã表ããã¯ãã«ã¨ãããã®ããwell-definedï¼æ°å¦ç¨èªï¼ã§ãªããã¨ãæ¬è³ªçãªåå ã§ããæ¬è³ªçã«åããã®ãå ¥åãã¦ããã®ã«ç°ãªãçµæãè¿ãã¨ããã¢ãã«ã¯ä¸å¤ã§ã¯ãªã (non-invariant) ã¨è¨ãã¾ããGNN ã®ææ°ååã§ãç´¹ä»ããããã«ãä¸å¤ãªã¢ãã«ãè¨è¨ãããã¨ã¯æ©æ¢°å¦ç¿ã®ç 究ãç¹ã«ã°ã©ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã§ã¯éè¦ãªãããã¯ã§ããæ¬ç¨¿ã®åé ã§è¿°ã¹ãå è«æ On the Expressive Power of Spectral Invariant Graph Neural Networks [Zhang+ ICML 2024] ã®ã¿ã¤ãã«ã«ãã invariant ã¨ããã®ã¯ãã®ãã¨ã§ããä¸å¤ (invariant) ãªãã¤ã¾ãåãé¨å空éãå ¥ãããåãçµæãè¿ã£ã¦ãããããªããã¥ã¼ã©ã«ãããã¯ã¼ã¯ããã³è¡¨ç¾ãè¦ã¤ãããã§ãã
ãã®è°è«ã¯é«æ¬¡å ã®å ´åãæç«ãã¾ããä¾ãã°ãæ£è¦ç´äº¤åºåºãé¨å空éã®è¡¨ç¾ã¨ãã¦ç¨ãããããããã¾ããããã㯠well-defined ã§ã¯ããã¾ãããä¸è¿°ã®ããã«ãåºåºã®ç¬¦å·ãå®ã¾ãã¾ãããããããé«æ¬¡å ã§ã¯åºåºãå転ããã¦ãåãé¨å空éã表ãã®ã§ãããåé¡ã¯è¤éã«ãªãã¾ãã
é¨å空éã¯å°å½±è¡åã§è¡¨ç¾ããã¨ãã
é¨å空éã¯å°å½±è¡åã§è¡¨ç¾ããã¨ããã§ããå°å½±è¡åã¯é¨å空éã«ã¨ã£ã¦ä¸æãªã®ã§ well-definedã§ãã
n 次å 空éä¸ã® d 次å é¨å空éã®æ£è¦ç´äº¤åºåºã並ã¹ãè¡åã ã¨ããã¨ãå°å½±è¡å㯠ã§ããè¡å ããã®é¨å空éã®è¯ãç¹å¾´éï¼è¡¨ç¾ï¼ã«ãªãã¾ãã
å°å½±è¡åã使ãã°ãã¾ãããçç±ãããããç´è¦³çã«è¿°ã¹ãã¨ã ã¯ãäºä¹ããã¦ããã ã§ããããã«ãäºä¹ãããã¨ã§ç¬¦å·ãçµ±ä¸ããã¾ãã
ä¾ãã°ãã ã®æ¹åãã ã¨ããã¨ãå°å½±è¡åã¯
ã§ããããã¯ãéæ¹åã®ãã¯ãã« ãç¨ãã¦è¨ç®ãã¦ãåãå¤ã«ãªãã¾ãã
é«æ¬¡å é¨å空éã«ããã¦å転ãããå ´åã«ããå転è¡åã¯è»¢ç½®ï¼éå転ï¼ããããã¨å ã«æ»ãã®ã§ããã¯ããäºä¹ãããã¨å転æåãçµ±ä¸ãããå°å½±è¡åã¯ä¸æã«ãªãã¾ãã
å¾æ®µã®ã¢ãã«ï¼å¤å±¤ãã¼ã»ãããã³ãªã©ï¼ã«æ¸¡ãã¨ãã«ã¯ãå°å½±è¡å ã®è¦ç´ ãä¸åã«ä¸¦ã¹ã 次å ãã¯ãã«ãç¨ããã°ããã§ãã
1 次å é¨å空éã表ç¾ããã¨ãããã¨ã㨠次å ã®è¡¨ç¾ãç¨ãã¦ããã®ã«ããããããã¯ããã«æ¬¡å ã®å¤§ãã 次å ã®ãã¯ãã«ã使ãã¨ãããã£ã¦èªç±åº¦ãå¢ãããã§ãããå®ã¯ããã§ã¯ãªããããã¯å¿ è¦ååãªèªç±åº¦ã§ãä¸æãªè¡¨ç¾ã«ãªã£ã¦ãã¾ããå°å½±è¡åã¯è¡¨é¢çã«ã¯ 次å ã§ãããã©ã³ã¯ãå¶éããã¦ããã®ã§èªç±åº¦ã¯ããã»ã©é«ãããã¾ããã1 次å é¨å空é㯠次å ãä¸è¬ã« d 次å é¨å空é㯠次å ã®å¤æ§ä½ããªããã¨ãç¥ããã¦ãã¾ãã
n 次å 空éä¸ã® d 次å é¨å空éï¼ ã©ã³ã¯ã ã®å°å½±è¡åï¼ã®éåã®ãã¨ãã°ã©ã¹ãã³å¤æ§ä½ (Grassmannian) ã¨ãããå¹¾ä½å¦ã®åéã§å¤ãããç 究ããã¦ãã¾ãããã詳細ãªå®ç¾©ãæ°å¦çãªæ§é ã«ã¤ãã¦ã¯ Grassmannian - Wikipedia ãªã©ãåç §ãã¦ãã ããã
On the Expressive Power of Spectral Invariant Graph Neural Networks [Zhang+ ICML 2024] ã§ã¯ãã°ã©ãã©ãã©ã·ã¢ã³ã®åºæ空éãå°å½±è¡å ã§è¡¨ãã ããã³ãã®åºæå¤ã辺 ã®ç¹å¾´éã¨ãã¦ç¨ããã°ã©ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã®ã¢ã¼ããã¯ãã£ãææ¡ãã¦ãã¾ãããã®æ¹æ³ã«ããããã°ã©ãã©ãã©ã·ã¢ã³ã®ä½åºæå¤ã®åºæ空éã=ãã°ã©ãä¸ã®æ»ãããªä¿¡å·ã表ãé¨å空éãã®æ å ±ãã°ã©ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã«åãå ¥ãããã¨ãã§ãã¾ããè«æã§ã¯ãã®æ¹å¼ãç¨ãããã¨ã§è¡¨ç¾è½åãåä¸ãããã¨ã示ããã¦ãã¾ãã
ãããã«
ä¸æã«è¡¨ç¾ã§ããã¨è²ã ã¨å¬ããã®ã§ãé¨å空éãè¨ç®æ©ä¸ã§æ±ãå¿ è¦ãçãããã¨ããããå°å½±è¡åãèãã¦ã¿ãã®ã¯è¯ãæ¦ç¥ã ã¨æãã¾ãã
ããæ½è±¡çãªæè¨ã¨ãã¦ã¯ã表ç¾ãã¢ãã«ã well-defined ã«ãªã£ã¦ããããä¸å¤ã«ãªã£ã¦ããããã¨ãããã¨ã¯ãçã®è¯ãã¢ãã«ãè¨è¨ããããã«å¸¸ã«æ°ã«ããã¦ããã¨è¯ãã§ãããã
æ¬ç¨¿ã§ã¯ãOn the Expressive Power of Spectral Invariant Graph Neural Networks [Zhang+ ICML 2024] ã®ç´°ããªã¨ããã¾ã§ã¯ç´¹ä»ã§ãã¾ããã§ãããããã°ã©ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ãã®ç¬¬ 6 ç« ã¨ç¬¬ 8 ç« ãèªãã°ãã®è«æãã¹ã ã¼ãºã«èªã¿é²ããããããã«ãªã£ã¦ããã¨æãã¾ãããã®è«æ㯠ICML 2020 ã®ååºè«æè³ (outstanding paper award) ãåè³ãã On Learning Sets of Symmetric Elements [Maron+ ICML 2020] ã®èè ã§ãã Haggai Maron ã®ã°ã«ã¼ãã®è«æã§ãå質ãé«ãã®ã§ãèå³ãæã£ãæ¹ã¯èªãã§ã¿ããã¨ããããããã¾ãã
æå¾ã«ãæ¬ç¨¿ã«èå³ãæã£ãæ¹ã¯ãã²ãã°ã©ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ããèªãã§ããã ããã¨å¬ããã§ãã
追è¨ï¼[Ishibashi+ NAACL 2024] ã [Kobayashi+ NAACL 2024] ã¨èª¤è¨ãã¦ããã®ãä¿®æ£ãã¾ããã失礼ãããã¾ããã
é£çµ¡å : @joisino_ / https://joisino.net