2021年，圖機器學習走勢會怎樣？

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"年底是總結和預測的好時機。2020 年圖機器學習（Graph ML）在機器學習領域大獲成功。本文中，我徵求了在圖機器學習和它的應用方面的著名研究者的意見，試圖總結出去年的一些亮點，並預測 2021 年的前景。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/47\/47e5ca0768673d4c01f21ec80994a8ac.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"italic"},{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"本文最初發表於 Towards Data Science 博客，經原作者 Michael Bronstein 授權，InfoQ 中文站翻譯並分享。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"超越消息傳遞"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/williamleif.github.io\/","title":null,"type":null},"content":[{"type":"text","text":"Will Hamilton"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，麥吉爾大學（McGill University）助理教授，Mila CIFAR 主席，"},{"type":"link","attrs":{"href":"http:\/\/snap.stanford.edu\/graphsage\/","title":null,"type":null},"content":[{"type":"text","text":"GraphSAGE"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 作者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在 2020 年時，圖機器學習領域面臨了消息傳遞模式的基本限制。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"其中包括所謂的“瓶頸”問題、過平滑的問題，以及在表達能力上的理論限制。從長遠來看，我希望到 2021 年，我們會發現圖機器學習的下一個大範式。雖然我不確定下一代圖機器學習算法將會是什麼樣的，但是我相信，要取得進展，就必須突破 2020 年及之前主導該領域的消息傳遞方案。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"我也希望 2021 年圖機器學習能夠進入更具影響力和挑戰性的應用領域。目前大量的研究集中在簡單、同質的節點分類任務上。同時，我也希望看到方法能夠發展到需要更復雜的算法推理的任務中，如涉及知識圖、強化學習和組合優化等。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"算法推理"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/5b\/5b49f4acf1006b5253f5321fe9a3f92b.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"指針圖網絡結合了經典計算機科學的結構歸納偏好。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/petar-v.com\/","title":null,"type":null},"content":[{"type":"text","text":"Petar Veličković"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，DeepMind 高級研究員，"},{"type":"link","attrs":{"href":"https:\/\/petar-v.com\/GAT\/","title":null,"type":null},"content":[{"type":"text","text":"Graph Attention Networks"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" （圖注意力網絡）作者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在機器學習領域，2020 年時，圖表示學習已經明顯和不可逆轉地成爲一流的公民。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"這一年有太多不能簡單列舉的進步，但是我個人最喜歡神經算法推理。神經網絡在插值方面傳統上非常強大，但衆所周知，它是一個很糟糕的外推器：因此也是不夠充分的推理器；因爲推理的一個主要特點是，它可以在分佈之外工作。推理任務很可能是進一步發展圖神經網絡的理想選擇，這不僅是因爲已知它們與此類任務非常吻合，而且還因爲許多現實世界的圖任務表現出同質性，也就是說，影響最大、可擴展性最好的方法，通常是更簡單的圖神經網絡形式。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"2020 年發表的多篇論文，以諸如神經圖靈機和可分化神經計算機等神經執行器 (neural executors）的歷史成功爲基礎，在如今普遍使用的圖機器學習工具的支持下，探索了神經執行器的理論極限，提出了一種新的、更強大的基於圖神經網絡的推理架構，並實現了神經推理任務的完美強泛化。儘管這類架構在 2021 年自然會成爲組合優化的勝利，但我個人最興奮的是，預訓練算法執行器能夠使我們把經典算法應用到那些過於原始或者不適合算法的輸入。舉例來說，我們的 XLVIN 代理就是利用這些概念，讓圖神經網絡在強化學習過程中執行一種價值迭代風格的算法，儘管底層 MDP 的具體內容還不清楚。我認爲，2021 年，圖神經網絡在一般強化學習中的應用時機將成熟。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"關係結構發現"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/43\/4317174d38beaab615692d11b43cfa8d.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"圖神經網絡可以學習一個狀態轉換圖（右），解釋一個複雜的多粒子系統（左）。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/tkipf.github.io\/","title":null,"type":null},"content":[{"type":"text","text":"Thomas Kipf"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，Google Brain（谷歌大腦）研究科學家。"},{"type":"link","attrs":{"href":"https:\/\/tkipf.github.io\/graph-convolutional-networks\/","title":null,"type":null},"content":[{"type":"text","text":"Graph Convolutional Networks"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"（圖卷積網絡）作者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"最近幾年，基於圖的神經網絡模型被廣泛採用以來，在圖機器學習領域出現了計算結構和數據結構分離的趨勢。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在最近一次 ICML "},{"type":"link","attrs":{"href":"https:\/\/slideslive.com\/38930558\/relational-structure-discovery","title":null,"type":null},"content":[{"type":"text","text":"研討會的演講"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"中，我將這種趨勢稱爲關係結構發現。典型地，我們設計圖神經網絡是爲了在數據集提供的固定（或時間上不斷演進）的結構上傳遞消息，也就是使用數據集的節點和邊作爲計算模型結構或消息傳遞結構的黃金標準。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在 2020 年，我們看到，人們對能夠調整計算結構的模型的興趣不斷上升，也就是它們使用什麼組件作爲節點，以及在哪些節點之間傳遞信息，而不僅僅是簡單的基於關注的模型。2020 年有影響力的例子包括 Amortised Causal Discovery，它利用神經關係推斷從時間序列數據中推斷（和推理）因果圖，具有可學習指針和關係機制的圖神經網絡、自適應計算圖的基於學習網格的物理模擬器，以及學習推斷在其上執行計算的抽象節點的模型。這一發展具有廣泛的意義，因爲它使我們能夠有效地利用 圖神經網絡架構在其他領域（如文本或視頻處理）中提供的對稱性（如節點置換同變性）和歸納偏好（如成對交互函數的建模）。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"放眼未來，我希望我們能夠看到，對於給定的數據和任務，在不依賴顯式監督的情況下，如何學習最佳的計算圖結構（包括節點和關係）方面有很多進展。通過檢驗這個學習到的結構，並通過推導一個學習到的模型來更好地解釋和解釋用於解決問題的計算，很可能使我們可以進一步類推因果推理。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"表達能力"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/haggaim.github.io\/","title":null,"type":null},"content":[{"type":"text","text":"Haggai Maron"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，英偉達研究科學家，《"},{"type":"link","attrs":{"href":"http:\/\/irregulardeep.org\/How-expressive-are-Invariant-Graph-Networks-(2-2","title":null,"type":null},"content":[{"type":"text","text":"可證明具有表達能力的高維圖形神經網絡"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"》（"},{"type":"text","marks":[{"type":"italic"},{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"provably expressive high-dimensional graph neural networks"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"）的作者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"很多優秀的論文都討論了不同類型的圖神經網絡架構的表達能力，並說明了其在受深度和寬度限制時的基本表達能力極限，描述了用圖神經網絡檢測和計數什麼樣的結構，說明了用固定數量的圖神經網絡來處理多個圖任務是沒有意義的，並且提出了一個迭代圖神經網絡，可以學習如何自適應地終止消息傳遞過程。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在 2021 年，我很樂意看到在圖的生成模型的原則性方法、圖與圖神經網絡的匹配和圖神經網絡的表達能力之間的聯繫、諸如圖象和音頻等結構化數據的學習、以及在神經網絡社區和處理場景圖的計算機視覺社區之間建立更緊密的聯繫方面取得進展。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"可擴展性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/rusty1s.github.io\/#\/","title":null,"type":null},"content":[{"type":"text","text":"Matthias Fey"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，多特蒙德工業大學（TU Dortmund）博士生，"},{"type":"link","attrs":{"href":"https:\/\/pytorch-geometric.readthedocs.io\/en\/latest\/","title":null,"type":null},"content":[{"type":"text","text":"PyTorch Geometric"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 和 "},{"type":"link","attrs":{"href":"https:\/\/ogb.stanford.edu\/","title":null,"type":null},"content":[{"type":"text","text":"Open Graph Benchmark"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 的開發者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"圖神經網絡的可擴展性問題是 2020 年圖機器學習研究的熱點問題之一。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"有些方法通過將預測和傳播進行解耦，從而簡化了底層計算。很多文章都將不可訓練的傳播模式和獨立於圖的模塊簡單地結合在一起作爲前處理或後處理步驟。這樣就會產生極好的運行時，並可以顯著地實現同構圖的平均性能。我很想知道，在訪問越來越大的數據集的過程中，我渴望看到怎樣才能進步，怎樣才能以一種可擴展的方式使用訓練和表達性傳播。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"動態圖"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/3b\/3b66dcb0b07591b93107f2b683e24a14.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/www.emanuelerossi.co.uk\/","title":null,"type":null},"content":[{"type":"text","text":"Emanuele Rossi"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，倫敦帝國學院博士生，Twitter 的機器學習研究員，"},{"type":"link","attrs":{"href":"https:\/\/towardsdatascience.com\/temporal-graph-networks-ab8f327f2efe","title":null,"type":null},"content":[{"type":"text","text":"Temporal Graph Networks"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"（時態圖網絡）的作者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"很多有趣的圖機器學習應用在本質上是動態的，圖的拓撲結構和屬性也是隨時間變化的。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"不管是社會網絡，金融交易網絡，還是用戶 - 商品互動網絡，都是如此。到目前爲止，大多數對圖機器學習的研究都集中於靜態圖。少量嘗試處理動態圖的工作主要是考慮了離散時間動態圖，也就是在規則間隔內的一系列圖的圖快照。到 2020 年，我們將看到關於連續時間動態圖的一系列新出現的更一般的研究成果，這些研究成果可被視爲定時事件的異步流。同時，動態圖模型的首個有趣的成功應用也開始出現：我們看到了虛假賬戶檢測、欺詐檢測，以及流行病控制的蔓延。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"我認爲，我們只是在這個振奮人心的方向上探索，許多有趣的問題還沒有得到解決。其中重要的開放性問題包括可擴展性、對動態模型的更好的理論理解，以及在單一框架中結合信息的空間和時間擴散。爲了確保能夠更好地評估和跟蹤進展情況，我們還需要更可靠和更具挑戰性的基準。最終，我希望看到更多動態圖神經架構的成功應用，尤其是在工業領域。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"新硬件"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/77\/77da8786f2d9869f81c34d8ea32d1607.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Graphcore 是一家爲圖機器學習開發新硬件的半導體公司。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Mark Saroufim，"},{"type":"link","attrs":{"href":"https:\/\/www.graphcore.ai\/","title":null,"type":null},"content":[{"type":"text","text":"Graphcore"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 機器學習工程師。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我想不出有哪個客戶沒有在生產中部署過圖形神經網絡，或者正在計劃部署。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"這一趨勢的部分原因是，在自然語言處理、蛋白質設計或分子特性預測等應用中，傳統上沒有采用自然圖結構，而是採用已有的成熟機器學習模型（如 Transformers）可以接受的數據序列。但是我們知道 Transformers "},{"type":"link","attrs":{"href":"https:\/\/thegradient.pub\/transformers-are-graph-neural-networks\/","title":null,"type":null},"content":[{"type":"text","text":"僅僅是圖神經網絡"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，它用來作爲鄰域聚集函數。在計算中，當某些算法勝出並不是因爲它們很適合解決某個問題，而是因爲它們在現有硬件上運行得很好時，這一現象被稱爲硬件彩票（Hardware Lottery）：在 GPU 上運行的 Transformers 就是這種情況。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在 Graphcore，我們建立了一個擁有 1472 個內核的新 MIMD 架構，能夠同時運行 8832 個程序，我們稱之爲智能處理單元（Intelligence Processing Unit ，IPU）。 結構非常適合於加速圖神經網絡。Poplar 軟件棧利用稀疏性的優勢，爲計算圖中的不同節點分配了不同的核心。在適合 IPU 的 900 MB 內存的模型中，我們的架構比 GPU 的吞吐量有很大的提升；否則，只需要幾行代碼，就可以將模型分佈在數千個 IPU 上。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"我很高興看到我們的客戶利用我們的架構建立了"},{"type":"link","attrs":{"href":"https:\/\/www.graphcore.ai\/resources\/research-papers","title":null,"type":null},"content":[{"type":"text","text":"大量的研究"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，包括 SLAM 的束調整（bundle adjustment）、使用局部更新訓練深度網絡或"},{"type":"link","attrs":{"href":"https:\/\/www.graphcore.ai\/mk2-benchmarks","title":null,"type":null},"content":[{"type":"text","text":"加速"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"粒子物理學中的各種問題等應用。我希望在 2021 年能看到更多的研究人員利用我們先進的機器學習硬件。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"在工業、物理、醫學等領域的應用"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/ac\/ac7c15a4a2611d2cb38826cb30bfd912.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"MagicLeap 的 SuperGlue 使用 圖神經網絡解決了一個經典的特徵匹配計算機視覺問題。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/ivanovml.com\/","title":null,"type":null},"content":[{"type":"text","text":"Sergey Ivanov"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，Criteo 研究科學家，《"},{"type":"link","attrs":{"href":"https:\/\/graphml.substack.com\/","title":null,"type":null},"content":[{"type":"text","text":"圖機器學習通訊"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"》（"},{"type":"text","marks":[{"type":"italic"},{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Graph Machine Learning newsletter"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"）編輯。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"對於圖機器學習研究來說，這是一個驚人的年份。大多數大型的機器學習會議都會有 10~20% 的關於這一領域的論文，每個人都會在這一範圍內找到自己喜歡的有趣的圖主題。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/gm-neurips-2020.github.io\/","title":null,"type":null},"content":[{"type":"text","text":"Google Graph Mining"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 團隊在 NeurIPS 上的表現非常突出。從 312 頁的"},{"type":"link","attrs":{"href":"https:\/\/gm-neurips-2020.github.io\/master-deck.pdf","title":null,"type":null},"content":[{"type":"text","text":"演講"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 來看，可以說谷歌在生產中使用圖形方面的比任何人都先進。他們的應用主要有新冠肺炎病毒的建模、欺詐檢測、隱私保護等，這些都是基於圖機器學習的解決方案。另外， DeepMind 也在其產品中推出了用於谷歌地圖全球範圍內的出行"},{"type":"link","attrs":{"href":"https:\/\/deepmind.com\/blog\/article\/traffic-prediction-with-advanced-graph-neural-networks","title":null,"type":null},"content":[{"type":"text","text":"時間預測"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"的圖神經網絡。這一方法的一個有趣的細節是，它集成了一個遷移學習模型，並將類似的採樣子圖選取到一個批次，用於圖神經網絡的參數訓練。這樣，先進的超參數調優就可以提高實時到達時間估計的正確性，達到最高 50%。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"另一個引人注目的應用是圖神經網絡，它由 MagicLeap 公司開發，該公司專注於 3D 計算機生成圖形。在 SuperGlue 的架構中，應用了圖神經網絡來進行圖像的特徵匹配：這是 3D 重建、位置識別、定位和映射的一個重要課題。端對端特徵表示與最佳的傳輸優化組合，在實時室內和室外姿勢估計方面都是領先的。這一結果只是 2020 年時所取得成就的一個縮影。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"明年，我相信我們將看到圖機器學習發展在"},{"type":"link","attrs":{"href":"https:\/\/medium.com\/criteo-labs\/top-applications-of-graph-neural-networks-2021-c06ec82bfc18","title":null,"type":null},"content":[{"type":"text","text":"工業領域"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"中的進一步使用。這將包括生產管道和框架、新的開源圖數據集，以及電子商務、工程設計和製藥行業大規模部署圖神經網絡。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/70\/70e7e593b59890149fdd80e9a2a04c07.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"以圖形表示的粒子噴流。圖神經網絡正在被用來探測粒子物理學中的事件。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"http:\/\/theoryandpractice.org\/","title":null,"type":null},"content":[{"type":"text","text":"Kyle Cranmer"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，紐約大學物理學教授，希格斯玻色子的發現者之一。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"圖機器學習最近兩年在物理學領域非常流行，這是一個驚人的現象。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"早期粒子物理的深度學習工作，常常要求把數據轉化爲圖像表示，以配合卷積神經網絡，這是不自然的，因爲我們的數據並非天生的網格，而圖表示則非常稀疏。圖是我們數據更自然的表示方式。LHC 上的研究人員正在努力將圖機器學習與實時數據處理系統結合起來，以每秒處理數十億次碰撞。目前正在進行的一項工作是將圖機器學習和實時數據採集系統集成到"},{"type":"link","attrs":{"href":"https:\/\/news.fnal.gov\/2020\/09\/the-next-big-thing-the-use-of-graph-neural-networks-to-discover-particles\/","title":null,"type":null},"content":[{"type":"text","text":"推理服務器"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 中，並在 FPGA 和其它特殊硬件上實現算法。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"2020 年圖機器學習的另一大亮點是，證明了其歸納偏好可以與符號法相匹配。舉例來說，我們利用圖神經網絡來學習如何預測各種動態系統，然後對沿邊緣發送的信息進行符號迴歸。不使用基準力學，我們不僅可以恢復這些動力學系統的基準力學定律，而且也可以提取方程。我們很驚訝地發現，通過提取的符號方程可以重新引入到圖神經網絡中，以代替原來學習的分量，從而得到更好的對分佈外數據的泛化。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/3f\/3f71a4a3c16846516eaaffcae99feed6.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"GNNs 可以利用種羣圖進行疾病分類。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"http:\/\/campar.in.tum.de\/Main\/AneesKazi","title":null,"type":null},"content":[{"type":"text","text":"Anees Kazi"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，慕尼黑工業大學博士生，多篇關於醫學成像中圖機器學習的論文作者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在醫療領域，圖機器學習改變了多模態數據的分析方式，它非常類似於專家們在臨牀常規中從各個維度觀察病人的情況。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"近年來，圖機器學習在醫學影像和醫學應用方面的相關研究出現了巨大的高峯，其中包括腦區域分割，利用 MRI\/fMRI 數據對腦結構進行疾病預測、藥物效應分析等。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"2020 年，在圖機器學習的主題中，有幾個在醫學領域非常突出。首先是潛圖學習（latent graph learning），由於經驗性地定義給定數據的圖是迄今爲止最佳結果的瓶頸，所以它現在被自動學習潛圖結構的方法所解決。其次是數據歸集（data imputation）， 因爲在醫學領域許多數據集中，缺乏數據是常態化問題，根據圖鄰域的關係，基於圖的方法已經有助於進行數據推斷。再次，關於機器學習模型的可解釋性問題，對臨牀和技術專家而言，重視機器學習模型的研究成果，並將其可靠地集成到 CADx 系統中，具有重要意義。在醫學領域，2020 年的另一大亮點當然是新冠肺炎疫情的爆發，圖機器學習方法被用來檢測新冠肺炎病毒。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在 2021 年，圖機器學習可以被用於進一步改善機器學習模型的可解釋性，從而更好地做出決策。根據觀察，圖機器學習方法對圖的結構仍然比較敏感，所以對圖干擾的健壯性和對抗性攻擊是一個重要研究課題。最終，將自監督學習和圖機器學習結合起來應用到醫學領域將會非常有趣。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/b1\/b11c1d6ffd95aeb790ddd082103d4554.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"使用幾何機器學習架構 MaSIF 設計的腫瘤靶點的不同蛋白結合劑。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/people.epfl.ch\/bruno.correia\/?lang=en","title":null,"type":null},"content":[{"type":"text","text":"Bruno Correia"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，瑞士洛桑聯邦理工學院（EPFL）助理教授，蛋白質設計和免疫工程實驗室的負責人，"},{"type":"link","attrs":{"href":"https:\/\/github.com\/LPDI-EPFL\/masif","title":null,"type":null},"content":[{"type":"text","text":"MaSIF"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 的開發者之一。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在 2020 年，蛋白質結構預測方面取得了令人振奮的進展，而蛋白質結構預測是生物信息學的關鍵問題。但是，這些分子表面所呈現的化學和幾何圖形對蛋白質的功能是至關重要的。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"幾十年來，人們一直在研究表面分子特徵，但是這給機器學習方法帶來了挑戰。在蛋白質建模領域，來自幾何深度學習領域的方法令人印象深刻，因爲它們具有處理不規則數據的能力，這些數據尤其適用於蛋白質表徵。在 MaSIF 中，在基於網格的分子表面表徵上使用幾何深度學習來學習模式，可以預測蛋白質與其他分子（蛋白質和代謝物）的相互作用，並且可以將對接計算速度提高几個數量級。這樣就可以促進更大規模的蛋白質相互作用網絡的預測工作。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"通過對 MaSIF 框架的進一步開發，我們能夠在任何時候生成我們的表面和化學特徵，從而避免了所有的預計算階段。我希望這一進展能在蛋白質和小分子設計方面產生改變，並有助於加速生物藥物的長期發展。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/c6\/c658109fb336be5dd73469c95f1501e3.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"圖神經網絡被用於 Decagon 的多藥性副作用預測。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/dbmi.hms.harvard.edu\/people\/marinka-zitnik","title":null,"type":null},"content":[{"type":"text","text":"Marinka Zitnik"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，哈佛醫學院（Harvard Medical School）生物醫學信息學助理教授，"},{"type":"link","attrs":{"href":"http:\/\/snap.stanford.edu\/decagon\/","title":null,"type":null},"content":[{"type":"text","text":"Decagon"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 作者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2020 年，圖機器學習將進入生命科學領域，這是令人興奮的。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在精心設計的基準數據集上，我們已經看到，圖神經網絡不僅比早期方法優越，而且還能爲開發新藥以幫助人類和在基本層面理解自然開闢途徑。其重點研究領域包括單細胞生物學、蛋白質與結構生物學、藥物發現與重組。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"幾個世紀以來，科學方法，即科學家用來系統而有邏輯地解釋自然世界的基本科學實踐，基本上沒有變化。到 2021 年，我希望我們能夠在利用圖機器學習改變這一現狀方面取得重大進展。要做到這一點，我認爲我們需要設計出能夠優化和操縱網絡系統並預測其行爲的方法，例如基因組學，即人體的自然實驗，是如何影響疾病環境中人類的特性的。這一方法需要結合擾動和干預數據（而不僅僅是從我們的世界獲取觀測數據）。此外，我希望我們可以發展出更多的方法來了解可操作的表徵，並且可以很容易地將其應用於科學中可操作的假設。這種方法可以讓我們在高風險的環境 （如化學測試、粒子物理、人類臨牀試驗）下作出決策，我們需要準確而可靠的預測，以便對其作出有意義的解釋。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"作者介紹："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"Michael Bronstein，倫敦帝國理工學院教授，Twitter 圖機器學習研究負責人，CETI 項目機器學習主管、研究員、教師、企業家和投資者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"原文鏈接："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"https:\/\/towardsdatascience.com\/predictions-and-hopes-for-graph-ml-in-2021-6af2121c3e3d"}]}]}