ਅ(x࣠) vs ༧ଌ(y࣠) by DimeNet (Klicpera et al, 2020) Dipole Moment Energy U HOMO LUMO Heat Capacity Enthalpy H Dipole Moment Energy U HOMO LUMO Heat Capacity Enthalpy H 堣唒㷕统ח״✮庠כׯ䔲ر٦ةך䲧ִ倯如痥דכ㣐ֹז〳腉䚍ָ֮
ਅ(x࣠) vs ༧ଌ(y࣠) by DimeNet (Klicpera et al, 2020) Dipole Moment Energy U HOMO LUMO Heat Capacity Enthalpy H Dipole Moment Energy U HOMO LUMO Heat Capacity Enthalpy H ⦓鸞ְז ⼧ⴓ鏩㺁דֹ✮庠铎䊴 ֿכذأزر٦ة 鎮箺儗ח 鋅ׇגזְر٦ة ך穠卓 堣唒㷕统ח״✮庠כׯ䔲ر٦ةך䲧ִ倯如痥דכ㣐ֹז〳腉䚍ָ֮
H H N O C C C C H H H H H GNN Layer Ԯ ԮMessage 縧䳔ず㢌ז乼⡲ ˖ OO-JOFBSזו 鴟暴䗙ꆀכⰩ㘗涸חכ ֿךԮד⢪ֲ ԯ • sum, mean or max • attentive pooling ԯAggregate 縧䳔♶㢌ז乼⡲ <latexit sha1_base64="rQGx6XMOvsCjXRp7b9gbmKjgm1M=">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</latexit> hi 0 @hi, M j2Ni (hi, hj, eij) 1 A 鵚⩸ַךˑ.FTTBHF1BTTJOH˒ד刿倜 <latexit sha1_base64="+I1ZH8a510AHL/VRK05INyOdDHc=">AAACi3ichVHLSsNAFL2Nr1qtrboR3BRLxVWZaFEpLooiuOzDPqAtJYnTdmheJGmhhv6ASzcu6kbBhfgBfoAbf8BFP0FcVnDjwps0IFqsN0zmzJl77pyZK+oyMy1CBj5uanpmds4/H1hYDC6FwssrBVNrGxLNS5qsGSVRMKnMVJq3mCXTkm5QQRFlWhRbR85+sUMNk2nqqdXVaVURGiqrM0mwkCpVRMVu9mqsFo6SOHEjMg54D0TBi7QWfoQKnIEGErRBAQoqWIhlEMDErww8ENCRq4KNnIGIufsUehBAbRuzKGYIyLbw38BV2WNVXDs1TVct4SkyDgOVEYiRF3JPhuSZPJBX8vlnLdut4Xjp4iyOtFSvhS7Wch//qhScLWh+qyZ6tqAO+65Xht51l3FuIY30nfOrYS6Zjdmb5Ja8of8bMiBPeAO18y7dZWi2P8GPiF7wxbBB/O92jIPCdpzfjScyiWjq0GuVH9ZhA7awH3uQghNIQ97twyX04ZoLcjtckjsYpXI+T7MKP4I7/gL6JJMZ</latexit> hi <latexit sha1_base64="iVE2zwSrY7uMFp5lULN8y4s7ZXg=">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</latexit> eij 갥挿暴䗙ꆀ 鴟暴䗙ꆀ Update ˖ OO-JOFBSזו 剣ぢ鴟♳ד暴䗙ꆀ刿倜׃גַ갥挿暴䗙ꆀפさ䧭ׅ غٔؒ٦ءّٝ %JSFDUFE.1// כָָֿ֮㛇劤䕎
LayerNorm ˖ ぐ갥挿ך暴䗙كؙزٕ刿倜ׅꥷח"UUFOUJPOⰅְ ˖ 5SBOTGPSNFSכزهٗآⵖ秈ךזְ(SBQI"UUFOUJPO/FUXPSL ("5 㢌珏הזׇ ˖ 鷞ח5SBOTGPSNFS㘗ך4FMG"UUFOUJPO(//חֿֿהדֹ Transformer GNN Layer ⡂גְ˘ Embedding + Pos Encoding A Generalization of Transformer Networks to Graphs Dwivedi & Bresson (2020) https://arxiv.org/abs/2012.09699 Do Transformers Really Perform Bad for Graph Representation? Ying et al (2021) https://arxiv.org/abs/2106.05234 Communicative Representation Learning on Attributed Molecular Graphs Song et al (2020) https://www.ijcai.org/proceedings/2020/0392.pdf Graph-BERT: Only Attention is Needed for Learning Graph Representations Zhang et al (2020) https://arxiv.org/abs/2001.05140 Ying et al (2021) ͷGraphormerKDDCup 2021ͷ Graph-levelλεΫͷ༏উϞσϧͰΘΕͨ ✲㷕统ָ⸬ַזְהׁ/-1װ $//♧䫛ַח鋅ִ$7㢌ꬠ׃גֹ 5SBOTGPSNFSכⴓ㶨ةأؙ㢌ִךַ
are “heavy” (non-hydrogen) atoms • the valence minus the number of hydrogens • the atomic number • the atomic mass • the atomic charge • the number of attached hydrogens • whether the atom is contained in at least one ring %BZMJHIU ⾱㶨♶㢌ꆀ • hydrogen-bond acceptor or not? • hydrogen-bond donor or not? • negatively ionizable or not? • positively ionizable or not? • aromatic or not? • halogen or not? Rogers and Hahn, JCIM (2005) https://doi.org/10.1021/ci100050t Faber et al, JCTC (2017) https://doi.org/10.1021/acs.jctc.7b00577 .1//ד欽ְ갥挿٥鴟暴䗙 鸬竲ꆀٓكٕ ؛ٌ؎ٝؿؓوذ؍ؙأך⾱㶨♶㢌ꆀ &$'1ך،ٕ؞ٔؤيדכ ⾱㶨♶㢌ꆀח鸬竲⦼Ⰵךכ ִ֮זַ 갥挿ٓكٕך䕵ⶴ
inducing (latent) programs, and reasoning with these programs is a key to solving artificial intelligence. Recently, there have been a lot of success stories in the deep learning community related to learning neural networks capable of using trainable memory abstractions. Neural Abstract Machines & Program Induction • Differentiable Neural Computers / Neural Turing Machines (Graves+ 2014) • Memory Networks (Weston+ 2014) • Pointer Networks (Vinyals+ 2015) • Neural Stacks (Grefenstette+ 2015, Joulin+ 2015) • Hierarchical Attentive Memory (Andrychowicz+ 2016) • Neural Program Interpreters (Reed+ 2016) • Neural Programmer (Neelakantan+ 2016) • DeepCoder (Balog+ 2016) : 䩛竲ֹ涸٥鎸〾涸乼⡲㷕统דֹفؚٗٓيה׃ג䪔ִ״ֲחזגֹ 僇爙涸ז⻉㷕濼陎 輐さ׃גְַֽ