Tweet
Tweet

Slide 1

Slide 1 text

戣䊛♧㷕 ⻌嵲麣㣐㷕⻉㷕⿾䘔ⶼ䧭灇瑔䬿挿 81**$3F%% ⴓ㶨ךؚٓؿ邌植ה堣唒㷕统 ͖͕ͨΘ ͍͕ͪ͘ 䎃剢傈 抟 傈劤⻉㷕⠓痥㔐$4+⻉㷕ؿؑأة
 ر٦ة؟؎ؒٝأך⚅歲׾ך׊ְג׫תׇ׿ַ https://itakigawa.github.io/

Slide 2

Slide 2 text

˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖ (SBQI/FVSBM/FUXPSLTהכ ˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ
 剑鵚ך鑧 鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ 劤傈ךⰻ㺁 (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ

Slide 3

Slide 3 text

˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖ (SBQI/FVSBM/FUXPSLTהכ ˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ
 剑鵚ך鑧 鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ 劤傈ךⰻ㺁 (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ

Slide 4

Slide 4 text

堣唒㷕统ך⳿殢 ... "ׁ׿ה#ׁ׿ךⱖ溪׾ⴓֽ׷فؚٗٓي׾⡲׶׋ְ "ׁ׿PS#ׁ׿ ⱖ溪 فؚٗٓي ➂꟦ז׵知⽃ח鍑ֽ׷ֽו䩛갫כ荈ⴓד׮铡僇דֹזְ

Slide 5

Slide 5 text

堣唒㷕统倜׃ְفؚٗٓىؚٝךػٓت؎ي ♧菙暟⡤钠陎 ˑ֮׶ָהֲ˒ J’aime la musique I love music 갈㡮钠陎 堣唒缺鏬 馄鍑⫷ ؜٦ي"* 堣唒㷕统㢳ꆀךⰅ⳿⸂ך鋅劤⢽ַ׵فؚٗٓي׾ 荈⹛涸ח⡲䧭ׅ׷ ְְַ־׿ז 䪮遭

Slide 6

Slide 6 text

➙傈ך⢽ⴓ㶨ך崞䚍װ暟䚍ך✮庠 24"32413 ⴓ㶨ך崞䚍٥暟䚍ך✮庠 CH 3 N N H N H H 3 C N 0.739 ꣖㹱慬䏝 H 3 C H 3 C NH O N O N O CH3 O N NH 2 O CH3 Br CH3 N H 3 C H N S N O CH3 N OH CH3 CH3 N N N CH3 H 3 C H2 N NH2 Ⰵ⳿⸂ך 鋅劤⢽

Slide 7

Slide 7 text

ⴓ㶨ךؚٓؿ邌植ה(SBQI/FVSBM/FUXPSLT (//T ⴓ㶨ך崞䚍٥暟䚍ך✮庠 CH 3 N N H N H H 3 C N 0.739 ꣖㹱慬䏝 葿ղװ׶倯֮׷ֽ׸ו➙傈כ(//׾ך׊ְג׫תׇ׿ַ ⴓ㶨圓鸡 ⴓ㶨鎸鶢㶨 )BOTDI谏歊岀זו ⴓ㶨ךؚٓؿ邌植 N O C C C C H H H H H 暴䗙كؙزٕ 害欽ך
 堣唒㷕统 (SBQI
 /FVSBM
 /FUXPSLT ⴓ㶨ؿ؍ٝؖ٦ فٔٝز
 &$'1岀זו ✮庠⦼

Slide 8

Slide 8 text

⹛堣(//כ䎢ְٌتٔذ؍׾窟♧涸ח䪔ֲֿהָדֹ׷ Brc1cncc(Br)c1 4.*-&4 ⻉㷕圓鸡䒭 甧⡤圓鸡 ꨵ㶨朐䡾 俑㶵⴨ ؚٓؿ
 杞纏 䎗⡦涸
 挿꧊さ 瑞꟦涸ⴓ䋒
 نُٔ٦ي 甧⡤ꂁ䏟 1VC$IFN%  ⽃穠さך
 㔐鯄ך荈歋䏝 O N N N H NH N N N CH3 CH3 䫑䝤䚍航泊⶝ $*% (//ד䪔ִ׷ٌتٔذ؍כ
 杞纏ךչؚٓؿպ״׶䎢ְ / #S #S

Slide 9

Slide 9 text

ⴓ㶨邌植 CC1CCNO1 如⯋ 如⯋ 如⯋ 宏稆 *NQMJDJU 宏稆 &YQMJDJU $BOPOJDBM4.*-&4 ⴓ㶨ךؚٓؿ邌植 N O C C C C H H H H H 暴䗙كؙزٕ .0-䕎䒭 ⾱㶨呌ךYZ[ ⻉㷕穠さ

Slide 10

Slide 10 text

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ ⴓ㶨$ / 0 'ך穈さַׇ׵䧭׷⾱㶨 )⟃㢩 תדךⴓ㶨 ך㸜㹀%圓鸡ה珏ך暟䚍⦼׾鎘皾 #-:1( EG Q MFWFM 2.%BUBTFU 3BNBLSJTIOBO Sci Data 1, 140022 (2014). https://doi.org/10.1038/sdata.2014.22

Slide 11

Slide 11 text

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ N O C C C C H H H H H 殢湡ך ֿךⴓ㶨ך NPM׾錁㻊

Slide 12

Slide 12 text

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ Dipole moment Isotropic polarizability Highest occupied molecular orbital (HOMO) energy Lowest unoccupied molecular orbital (LUMO) energy Gap between HOMO and LUMO Electronic spatial extent Zero point vibrational energy Internal energy at 0K Internal energy at 298.15K Enthalpy at 298.15K Free energy at 298.15K Heat capavity at 298.15K Atomization energy at 0K Atomization energy at 298.15K Atomization enthalpy at 298.15K Atomization free energy at 298.15K Rotational constant A Rotational constant B Rotational constant C N O C C C C H H H H H ֿ׸׵ך%'5鎘皾⦼׾
 %'5׃זְד ✮庠ׅ׷ةأؙ ةذח邌爙

Slide 13

Slide 13 text

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ node[0] node[1] Edge 0 Edge 4 鴟ךぢֹ׮ 罋䣁דֹ׷ 0 1

Slide 14

Slide 14 text

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ 갥挿ך暴䗙كؙزٕ 如⯋ 鴟ך暴䗙كؙزٕ 如⯋ 갥挿ך⡘縧كؙزٕ 如⯋ 갥挿ך⾱㶨殢〾 (one-hot
 encoding) Single Double Triple Aromatic

Slide 15

Slide 15 text

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ 갥挿ך暴䗙كؙزٕ 如⯋ 鴟ך暴䗙كؙزٕ 如⯋ 갥挿ך⡘縧كؙزٕ 如⯋ 갥挿ך⾱㶨殢〾 (one-hot
 encoding) Single Double Triple Aromatic

Slide 16

Slide 16 text

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ 갥挿ך暴䗙كؙزٕ 如⯋ H C N O F (one-hot
 encoding) 0 1 2 3 4 5 6 7 8 9 10 https://pytorch-geometric.readthedocs.io/en/latest/_modules/torch_geometric/datasets/qm9.html ⾱㶨殢〾 BSPNBUJD TQ TQ TQ 穠さׅ׷)ך侧 ֿ׸⟃㢩ח׮ְ׹ְ׹剣׶ִ׷ֿהח岣䠐 N O C C C C H H H H H

Slide 17

Slide 17 text

(//ָװ׷ֿה 갥挿ך暴䗙كؙزٕ 如⯋ NPMY NPMQPT 갥挿ך⡘縧كؙزٕ 如⯋ NPMFEHF@BUUS NPMFEHF@JOEFY 鴟ך暴䗙كؙزٕ 如⯋ 鴟ה穠さ⾱㶨ل،ך㼎䘔 ⴓ㶨ךؚٓؿ邌植 N O C C C C H H H H H 暴䗙كؙزٕ NPMZ %'5鎘皾⦼ 如⯋ 㔿㹀؟؎ؤך كؙزٕ邌植 3FHSFTTPS
 )FBE
 .-1זו (// 〳㢌؟؎ؤךⴓ㶨ؚٓؿַ׵չ㔿㹀؟؎ؤךكؙزٕ邌植պפך㢌䳔׾㷕统

Slide 18

Slide 18 text

呌הז׷䪮遭♳ך銲⟝ (// 麩ֲ؟؎ؤךⴓ㶨ؚٓؿ׾Ⰵ⸂׃ג׮
 ְא׮չ㔿㹀؟؎ؤךكؙزٕ邌植պָ⳿⸂ׁ׸זְהְֽזְ CH 3 N N H N H H 3 C N H 3 C H 3 C NH O N O N O CH3 O N NH 2 O CH3 Br CH3 N H 3 C H N S N O CH3 N OH CH3 CH3 N N N CH3 H 3 C H2 N NH2 (// (// (// (// 갥挿侧װ 鴟侧ָ殯ז׷ ⴓ㶨ؚٓؿ ְא׮ ずׄ؟؎ؤך كؙزٕ׾⳿⸂

Slide 19

Slide 19 text

˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖ (SBQI/FVSBM/FUXPSLTהכ ˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ
 剑鵚ך鑧 鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ 劤傈ךⰻ㺁 (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ

Slide 20

Slide 20 text

堣唒㷕统ךٌرٕ 堣唒㷕统ٌرٕ 堣唒㷕统㢳ꆀךⰅ⳿⸂ך鋅劤⢽ַ׵فؚٗٓي׾ 荈⹛涸ח⡲䧭ׅ׷ ְְַ־׿ז 䪮遭 㢳侧ךػًٓةך⦼׾⹛ַ׃ג ⳿⸂ָ劄׬⦼חז׷״ֲחׅ׷

Slide 21

Slide 21 text

ٌرٕך㷕统ػًٓةך⦼׾剑黝חׅ׷ ✮庠⦼ 侄䌌ך⦼ 鋅劤⢽׾ ✮庠 ✮庠⦼ה侄䌌ך⦼ָ鵚בֻ״ֲח ػًٓة׾锃侭

Slide 22

Slide 22 text

ٌرٕך㷕统ػًٓةך⦼׾剑黝חׅ׷ O⦐ך鋅劤ךⰋ⡤חאְג
 ✮庠⦼ה侄䌌ך⦼ָ鵚בֻ״ֲח ػًٓة׾锃侭

Slide 23

Slide 23 text

ٌرٕך㷕统ػًٓةך⦼׾剑黝חׅ׷ O⦐ך鋅劤ךⰋ⡤חאְג
 ✮庠⦼ה侄䌌ך⦼ָ鵚בֻ״ֲח ػًٓة׾锃侭 ꟼ侧խխխխך⦼ָ㼭ֻׁז׷ ״ֲחػًٓة׾锃侭

Slide 24

Slide 24 text

ٌرٕך㷕统ػًٓةך⦼׾剑黝חׅ׷ O⦐ך鋅劤ךⰋ⡤חאְג
 ✮庠⦼ה侄䌌ך⦼ָ鵚בֻ״ֲח ػًٓة׾锃侭 ꟼ侧խխխխך⦼ָ㼭ֻׁז׷ ״ֲחػًٓة׾锃侭 ˖ խ׾ٓٝتي⦼דⴱ劍⻉ ˖ խխխך⦼ָ㼭ֻׁז׷״ֲחխ׾׍׳׏ה׌ֽ㢌ִ׷ ػًٓة锃侭ךװ׶ַ׋ ֻ׶ִַ׃ 㷕统桦
 TUFQTJ[F

Slide 25

Slide 25 text

⺟ꂁ꣬♴岀 (SBEJFOU%FTDFOU ˖ խխխך⦼ָ㼭ֻׁז׷״ֲחխ׾׍׳׏ה׌ֽ㢌ִ׷ ⺟ꂁكؙزٕח嫰⢽ׅ׷ꆀ׾׍׳׏ה䒷ֻה
 剑㺔׶ך噰㼭⦼חぢֲַֿהָ׻ַ׏גְ׷ ⺟ꂁكؙزٕ ⨉䗍ⴓ խװ׾ ך㢌⻉ꆀ 㼰׃⹛ַ׃׋הֹך

Slide 26

Slide 26 text

鎘皾ؚٓؿה荈⹛䗍ⴓ醱꧟זٌرٕך⺟ꂁך鎘皾 堣唒㷕统ٌرٕ 㼭ֻׁ׃׋ְ㽯䏝 ⺟ꂁكؙزٕ ֿ׸׾וֲ鎘皾ׅ׷ַ הְֲ鑧

Slide 27

Slide 27 text

鎘皾ؚٓؿ稆怴皾ךさ䧭ꟼ侧ךؚٓؿ邌植 ⢽ ✮庠ٌرٕ 暴ח䠐㄂כזְ5PZ&YBNQMF

Slide 28

Slide 28 text

⢽ ✮庠ٌرٕ 暴ח䠐㄂כזְ5PZ&YBNQMF さ䧭ꟼ侧 鎘皾ؚٓؿ稆怴皾ךさ䧭ꟼ侧ךؚٓؿ邌植

Slide 29

Slide 29 text

⢽ ✮庠ٌرٕ 暴ח䠐㄂כזְ5PZ&YBNQMF さ䧭ꟼ侧 鎘皾ؚٓؿ稆怴皾ךさ䧭ꟼ侧ךؚٓؿ邌植

Slide 30

Slide 30 text

⢽ ✮庠ٌرٕ 暴ח䠐㄂כזְ5PZ&YBNQMF さ䧭ꟼ侧 鎘皾ؚٓؿ稆怴皾ךさ䧭ꟼ侧ךؚٓؿ邌植

Slide 31

Slide 31 text

⢽ ✮庠ٌرٕ 暴ח䠐㄂כזְ5PZ&YBNQMF さ䧭ꟼ侧 鎘皾ؚٓؿ稆怴皾ךさ䧭ꟼ侧ךؚٓؿ邌植

Slide 32

Slide 32 text

⢽ ✮庠ٌرٕ 暴ח䠐㄂כזְ5PZ&YBNQMF さ䧭ꟼ侧 鎘皾ؚٓؿ稆怴皾ךさ䧭ꟼ侧ךؚٓؿ邌植

Slide 33

Slide 33 text

⢽ ✮庠ٌرٕ 暴ח䠐㄂כזְ5PZ&YBNQMF さ䧭ꟼ侧 鎘皾ؚٓؿ稆怴皾ךさ䧭ꟼ侧ךؚٓؿ邌植

Slide 34

Slide 34 text

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ

Slide 35

Slide 35 text

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 1.4

Slide 36

Slide 36 text

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 0.17 1.4

Slide 37

Slide 37 text

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 0.17 0.05 1.4

Slide 38

Slide 38 text

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 0.17 0.05 1.0 1.4

Slide 39

Slide 39 text

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 0.17 0.05 1.0 0.95 1.4

Slide 40

Slide 40 text

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 0.17 0.05 1.0 0.95 8.7 1.4

Slide 41

Slide 41 text

GPSXBSE 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 0.17 0.05 1.0 0.95 8.7 1.4

Slide 42

Slide 42 text

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

Slide 43

Slide 43 text

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 1 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

Slide 44

Slide 44 text

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 1 -0.99 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

Slide 45

Slide 45 text

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 1 -0.99 0.3 0.17 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

Slide 46

Slide 46 text

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 1 -0.99 0.3 0.17 5 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

Slide 47

Slide 47 text

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 1 -0.99 0.3 0.17 5 -1 1 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

Slide 48

Slide 48 text

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 1 -0.99 0.3 0.17 5 -1 1 5.9 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

Slide 49

Slide 49 text

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾 さ䧭ח⢪ִ׷稆怴皾כ✮׭寸׭ג׉׸׊׸ך䗍ⴓⰕ䒭׮䭯׏גֶֻ ˖ ⸇幾⛦ꤐ
 UPSDIBEE UPSDITVC UPSDINVM UPSDIEJW ˖ ♲錬ꟼ侧
 UPSDITJO UPSDIDPT UPSDIUBO UPSDIBTJO  ˖ 䭷侧٥㼎侧٥ץֹ
 UPSDIQPX UPSDIMPH UPSDITRVBSF UPSDITRSU ˖ 遤⴨怴皾
 UPSDINBUNVM UPSDIEPU UPSDIJOWFSTF UPSDIEFU ˖ 乼⡲
 UPSDITVN UPSDINFBO UPSDINBY UPSDIBSHNJO 㢌侧ךさ䧭ח⢪ִ׷稆怴皾ך⢽ https://pytorch.org/docs/stable/torch.html#math-operations

Slide 50

Slide 50 text

GPSXBSE https://pytorch.org/tutorials/beginner/nlp/pytorch_tutorial.html 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4

Slide 51

Slide 51 text

劤䔲ח鎘皾׃׋ְךכ⺟ꂁ׾鎘皾ׅ׷㢌侧חꟼׅ׷⨉䗍ⴓ ⺟ꂁ كؙزٕ 荈⹛䗍ⴓ ٔغ٦أٌ٦س #BDL1SPQBHBUJPO 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 ׾㼰׃⹛ַ׃׋הֹך ך㢌⻉ꆀ ⺟ꂁ岀ָ⢪ְ׋ְ

Slide 52

Slide 52 text

CBDLXBSE さ䧭ꟼ侧ך䗍ⴓך鸬ꓲ䖒״׶ծ鎘皾ؚٓؿ׾⢪׏ג知⽃ח鎘皾דֹ׷ 5.9 5.9 5.9 1 5 -1 0.17 -1 0.3 -0.99 1 29.5 -1.00 1.75 1 -0.99 0.3 0.17 5 -1 1 5.9 = = = × × × × × × × ×

Slide 53

Slide 53 text

فؚٗٓيהכ瑔噰涸חכ稆怴皾ך⡦׵ַךさ䧭ꟼ侧 GPSXBSE CBDLXBSE 荈⹛䗍ⴓ 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 1 -0.99 0.3 0.17 5 -1 1 5.9 ✮庠⦼ MPTT⦼ ⺟ꂁكؙزٕ

Slide 54

Slide 54 text

⺟ꂁكؙزٕ 帾㾴㷕统ך؝؝ٗ䗍ⴓ〳腉זفؚٗٓىؚٝ ו׿זفؚٗٓي׮稆怴皾ךさ䧭ꟼ侧 鎘皾ؚٓؿ חז׏גִׁ ְ׸ל荈⹛䗍ⴓ⺟ꂁ꣬♴岀דػًٓة׾剑黝חדֹ׷ 㷕统桦
 TUFQTJ[F 荈⹛䗍ⴓד鎘皾 فؚٗٓيך♶然㹀ז皘䨽כػًٓةח׃גֶֹ䖓ד堣唒㷕统

Slide 55

Slide 55 text

劤傈ךⰻ㺁 ˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖ (SBQI/FVSBM/FUXPSLTהכ ˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ
 剑鵚ך鑧 鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ

Slide 56

Slide 56 text

(//♴鎸ך㢌䳔׾⡦׵ַךさ䧭ꟼ侧דرؠ؎ٝ׃׋׮ך 갥挿ך暴䗙كؙزٕ 如⯋ NPMY NPMQPT 갥挿ך⡘縧كؙزٕ 如⯋ NPMFEHF@BUUS NPMFEHF@JOEFY 鴟ך暴䗙كؙزٕ 如⯋ 鴟ה穠さ⾱㶨ل،ך㼎䘔 ⴓ㶨ךؚٓؿ邌植 N O C C C C H H H H H 暴䗙كؙزٕ NPMZ %'5鎘皾⦼ 如⯋ 㔿㹀؟؎ؤך كؙزٕ邌植 3FHSFTTPS
 )FBE
 .-1זו (// 〳㢌؟؎ؤךⴓ㶨ؚٓؿַ׵չ㔿㹀؟؎ؤךكؙزٕ邌植պפך㢌䳔׾㷕统

Slide 57

Slide 57 text

㛇劤麣Ⱗ.-1 㢳㾴ػ٦إفزٗٝ x =  x1 x2 x1 x2 1 1 w0 ji w00 i wkj 1 y import torch.nn as nn nn.Sequential(nn.Linear(2, 3), nn.ReLU(), nn.Linear(3, 2), nn.ReLu(), nn.Linear(2, 1)) UPSDIOO-JOFBS ˖ JO@GFBUVSFT ˖ PVU@GFBUVSFT ˖ CJBT 5SVF'BMTF Ⰵ⸂؟؎ؤ ⳿⸂؟؎ؤ CJBTך剣搀 1 Ⰵ⸂
 ؟؎ؤ CJBT ⳿⸂
 ؟؎ؤ

Slide 58

Slide 58 text

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO N O C C C C H H H H H N O C C C C H H H H H ぐ暴䗙كؙزٕך刿倜 MLPͳͲ (sum, mean or max) 갫殢ח⣛㶷׃זְ꧊秈乼⡲ MLPͳͲ UPSDI@HFPNFUSJDOO .FTTBHF1BTTJOH Ԯ ԯ ԰ Ԯ ԯ ԰

Slide 59

Slide 59 text

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO 1 2 3 4 5 6 0 1 0 0 0 0 1 0 1 1 1 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 0 1 0 0 0 0 1 0 1 1 1 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 0 0 1 0 1 1 1 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 0 0 1 0 1 1 1 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 = = = 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 3 4 5 2 + + + ꦄ䱸遤⴨זו׾⢪׏ג׮ ꧊秈ׅ׷乼⡲כ剅ֽ׷

Slide 60

Slide 60 text

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO ˖ UPSDI@HFPNFUSJDOO($/$POW ,JQGBOE8FMMJOH *$.- https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html ˖ UPSDI@HFPNFUSJDOO4"(&$POW )BNJMUPOFUBM /*14 PS ˖ UPSDI@HFPNFUSJDOO(SBQI$POW .PSSJTFUBM """* ˖ UPSDI@HFPNFUSJDOO(*/$POW 9VFUBM *$-3

Slide 61

Slide 61 text

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO ˖ UPSDI@HFPNFUSJDOO4DI/FU 4DI»UUFUBM /*14 ˖ UPSDI@HFPNFUSJDOO("5$POW 7FMJčLPWJćFUBM *$-3 https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html YZ[דך騃ꨄ׾鴟ꅾ׫ח׃ג꧊秈儗ח崞欽ׅ׷ ꧊秈乼⡲כءٝفٕ׌ָ،ذٝءّٝ堣圓דꅾ׫׾黝䘔涸ח 銲稆琎 ]]DPODBU

Slide 62

Slide 62 text

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO ˖ UPSDI@HFPNFUSJDOO%JNF/FU ,MJDQFSBFUBM *$-3 YZ[ח㼎׃ג⻉㷕涸זسً؎ٝ濼陎׾崞欽׃ג乼⡲׾⡲׶鴥׫תֻ׏׋⢽

Slide 63

Slide 63 text

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO 剑ⴱח⢽ח⳿׃׋2.ر٦ةדכ䖞勻䩛岀׾⮚馉ׅ׷✮庠礵䏝

Slide 64

Slide 64 text

3FBEPVUؚٓؿח㼎ׅ׷㔿㹀؟؎ؤ邌植׾䖤׷ N O C C C C H H H H H 3FBEPVU ˖ UPSDI@HFPNFUSJDOO(MPCBM"UUFOUJPO 㾴侧 ։ ˖ UPSDI@HFPNFUSJDOO4FU4FU ˖ UPSDI@HFPNFUSJDOOHMPCBM@BEE@QPPM
 UPSDI@HFPNFUSJDOOHMPCBM@NBY@QPPM
 UPSDI@HFPNFUSJDOOHMPCBM@NFBO@QPPM ˖ UPSDI@HFPNFUSJDOOHMPCBM@TPSU@QPPM (sum, mean or max) 갫殢ח⣛㶷׃זְ꧊秈乼⡲ 剑䖓ך銲稆דTPSU׃גزحفL׾DPODBU

Slide 65

Slide 65 text

(//♴鎸ך㢌䳔׾⡦׵ַךさ䧭ꟼ侧דرؠ؎ٝ׃׋׮ך 갥挿ך暴䗙كؙزٕ 如⯋ NPMY NPMQPT 갥挿ך⡘縧كؙزٕ 如⯋ NPMFEHF@BUUS NPMFEHF@JOEFY 鴟ך暴䗙كؙزٕ 如⯋ 鴟ה穠さ⾱㶨ل،ך㼎䘔 ⴓ㶨ךؚٓؿ邌植 N O C C C C H H H H H 暴䗙كؙزٕ NPMZ %'5鎘皾⦼ 如⯋ 㔿㹀؟؎ؤך كؙزٕ邌植 3FHSFTTPS
 )FBE
 .-1זו (// 〳㢌؟؎ؤךⴓ㶨ؚٓؿַ׵չ㔿㹀؟؎ؤךكؙزٕ邌植պפך㢌䳔׾㷕统

Slide 66

Slide 66 text

1Z5PSDI(FPNFUSJDך⢽ (PPHMF$PMBC

Slide 67

Slide 67 text

ٌرٕהGPSXBSE㹀纏 ⳿⸂׾鎘皾ׅ׷GPSXBSE 3FBEPVU .FTTBHF 1BTTJOH Y㾴 3FHSFTTPS )FBE .-1

Slide 68

Slide 68 text

㷕统ٌرٕךػًٓة׾剑黝חׅ׷ ؚٓؿ׾CBUDI@TJ[F ֿ באGFFE ٌرٕחⰅ׸ג✮庠׾䖤׷ MPTT׾鎘皾׃CBDLXBSE ⺟ꂁ꣬♴ ػًٓة刿倜 ה׶ִ֮׆FQPDI 侧涰䗳銲

Slide 69

Slide 69 text

岣䠐 ˖ 2.ךةأؙך㜥さכקרYZ[ YQPT ך䞔㜠ָ䗳銲׌ָծ
 ⽃秪ז(//כYFEHF@BUUS׮⢪׻זְךד黝㹅何㢌ָ䗳銲 ˖ ⢽ִלYFEHF@BUUSח⾱㶨꟦騃ꨄ׾⸇ִג$POWד崞欽 ˖ ׉ךתת⢪ֲ㜥さכ4DI/FUװ%JNF/FUָぢְגְ׷ ˖ Ⱅ䒭؝٦س؟ٝفٕ׮ثؑحؙ
 IUUQTHJUIVCDPNSVTUZTQZUPSDI@HFPNFUSJDUSFFNBTUFSFYBNQMFT

Slide 70

Slide 70 text

0QFO(SBQI#FODINBSL https://ogb.stanford.edu

Slide 71

Slide 71 text

劤傈ךⰻ㺁 ˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖ (SBQI/FVSBM/FUXPSLTהכ ˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ
 剑鵚ך鑧 鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ

Slide 72

Slide 72 text

(SBQI/FVSBM/FUXPSLT׾濼׷׋׭ךאךا٦أ https://ai.tencent.com/ailab/ml/KDD-Deep-Graph-Learning.html • KDD2020 Tutorial
 Deep Graph Learning: Foundations, Advances and Applications http://cse.msu.edu/~mayao4/dlg_book/index.html • Book "Deep Learning on Graphs" • CS224W: Machine Learning with Graphs (Stanford) http://web.stanford.edu/class/cs224w/ https://youtu.be/-UjytpbqX4A

Slide 73

Slide 73 text

1ZUPSDI(FPNFUSJD׾⢪׏׋عٝؤؔٝ $48 https://colab.research.google.com/drive/1DIQm9rOx2mT1bZETEeVUThxcrP1RKqAn

Slide 74

Slide 74 text

,BHHMF$)".14 YZ[ָ銲׷ةأؙ https://www.kaggle.com/c/champs-scalar-coupling/discussion/93972

Slide 75

Slide 75 text

*$-3"'BJS$PNQBSJTPOTPG(//T https://github.com/diningphil/gnn-comparison

Slide 76

Slide 76 text

劤傈ךⰻ㺁 ˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖ (SBQI/FVSBM/FUXPSLTהכ ˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ
 剑鵚ך鑧 鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ

Slide 77

Slide 77 text

(//כ涪㾜鷿♳ך䪮遭ד植㖈鹌遤䕎ד灇瑔ׁ׸גְ׷ ˖ ✲⵸㷕统ָדֹזְ ˖ %FFQחדֹזְ ˖ 㷕统דֹ׷邌植חꣲ歲ָ֮׷ Generalization and Representational Limits of Graph Neural Networks Garg, Jegelka, Jaakkola (ICML2020) PairNorm: Tackling Oversmoothing in GNNs Zhao, Akoglu (ICLR2020) Towards Deeper Graph Neural Networks with Differentiable Group Normalization Zhou, Huang, Li, Zha, Chen, Hu (NeurIPS2020) Strategies for Pre-training Graph Neural Networks Hu, Liu, Gomes, Zitnik, Liang, Pande, Leskovec (ICLR2020) GROVER: Self-Supervised Message Passing Transformer on Large-scale Molecular Graphs Rong, Bian, Xu, Xie, Wei, Huang, Huang (NeurIPS2020) DeepGCNs: Can GCNs Go as Deep as CNNs? Li, Müller, Thabet, Ghanem (ICCV2019)

Slide 78

Slide 78 text

➙傈ך鑧ך䘔欽٥涪㾜 ˖ ⻉㷕⿾䘔ך✮庠٥鷞さ䧭 ˖ ⴓ㶨ך欰䧭 ˖ (SBQI1PPMJOH Path Integral Based Convolution and Pooling for Graph Neural Networks Ma, Xuan, Guang Wang, Li, Liò (NeurIPS2020) Hierarchical Generation of Molecular Graphs using Structural Motifs Jin, Barzilay, Jaakkola (ICML2020) GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation Shi, Xu, Zhu, Zhang, Zhang, Tang (ICLR2020) RetroXpert: Decompose Retrosynthesis Prediction like A Chemist Yan, Ding, Zhao, Zheng, Yang, Yu, Huang (NeurIPS2020) A Graph to Graphs Framework for Retrosynthesis Prediction Shi, Xu, Guo, Zhang, Tang (ICML2020) Rethinking Pooling in Graph Neural Networks Mesquita, Souza, Kaski (NeurIPS2020)

Slide 79

Slide 79 text

1VSFMZ%BUB%SJWFOךꣲ歲WT$IFNBUJDB

Slide 80

Slide 80 text

תה׭ ˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖ (SBQI/FVSBM/FUXPSLTהכ ˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ
 剑鵚ך鑧 鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ