本週,機器學習和計算神經科學領域的頂級大會第 30 屆國際神經資訊處理系統大會(NIPS2016)在巴塞羅那舉辦,內容包括演講、展示和宣講和海報展示,在這裡可以一睹最新的機器學習研究。谷歌帶著 280 名員工強勢亮相,除了技術演講和海報展示外,他們還將舉辦研討會和多個 tutorials。
谷歌的研究一直走在機器學習的前沿,積極探索機器學習的各個方面,包括經典演算法以及像深度學習這樣的前沿技術,既關注理論也重視應用。他們在語言理解、語音、翻譯、視覺處理、排名和預測上的很多成果都依賴於機器智慧。在所有的任務中,他們收集了大量直接或間接的利益關係的證據,並開發學習理解和泛化的方法。
Invited Talk
標題:Dynamic Legged Robots
作者:Marc Raibert
新一代高效能的機器人正在離開實驗室進入現實世界,出現在辦公室、家庭以及一些普通機器無法到達的地方。這些新興機器人使用探測器來觀察周邊,並依靠其在環境中導航,理解環境,與環境互動。它們敏捷、靈巧和自主和智慧都在按照將人類從各種任務中解放出來的願景在不斷髮展進化。
28篇論文
1.論文:Boosting with Abstention
作者:Corinna Cortes, Giulia DeSalvo, Mehryar Mohri
論文地址:http://papers.nips.cc/paper/6335-boosting-with-abstention
2.論文: Community Detection on Evolving Graphs
作者:Stefano Leonardi, Aris Anagnostopoulos, Jakub Łącki, Silvio Lattanzi, Mohammad Mahdian
論文地址:http://papers.nips.cc/paper/6173-community-detection-on-evolving-graphs.pdf
3.論文:Linear Relaxations for Finding Diverse Elements in Metric Spaces
作者:Aditya Bhaskara, Mehrdad Ghadiri, Vahab Mirrokni, Ola Svensson
論文地址:http://papers.nips.cc/paper/6500-linear-relaxations-for-finding-diverse-elements-in-metric-spaces.pdf
4.論文:Nearly Isometric Embedding by Relaxation
作者:James McQueen, Marina Meila, Dominique Joncas
論文地址:http://papers.nips.cc/paper/6535-nearly-isometric-embedding-by-relaxation.pdf
5.論文:Optimistic Bandit Convex Optimization
作者:Mehryar Mohri, Scott Yang
論文地址:http://papers.nips.cc/paper/6429-optimistic-bandit-convex-optimization.pdf
6.論文:Reward Augmented Maximum Likelihood for Neural Structured Prediction
作者:Mohammad Norouzi, Samy Bengio, Zhifeng Chen, Navdeep Jaitly, Mike Schuster, Yonghui Wu, Dale Schuurmans
論文地址:http://papers.nips.cc/paper/6547-reward-augmented-maximum-likelihood-for-neural-structured-prediction.pdf
7.論文:Stochastic Gradient MCMC with Stale Gradients
作者:Changyou Chen, Nan Ding, Chunyuan Li, Yizhe Zhang, Lawrence Carin
論文地址:http://papers.nips.cc/paper/6359-stochastic-gradient-mcmc-with-stale-gradients.pdf
8.論文:Unsupervised Learning for Physical Interaction through Video Prediction
作者:Chelsea Finn*, Ian Goodfellow, Sergey Levine
論文地址:http://papers.nips.cc/paper/6161-unsupervised-learning-for-physical-interaction-through-video-prediction.pdf
9.論文:Using Fast Weights to Attend to the Recent Past
作者:Jimmy Ba, Geoffrey Hinton, Volodymyr Mnih, Joel Leibo, Catalin Ionescu
論文地址:http://papers.nips.cc/paper/6057-using-fast-weights-to-attend-to-the-recent-past.pdf
10.論文:A Credit Assignment Compiler for Joint Prediction
作者:Kai-Wei Chang, He He, Stephane Ross, Hal III
論文地址:http://papers.nips.cc/paper/6256-a-credit-assignment-compiler-for-joint-prediction.pdf
11.論文:A Neural Transducer
作者:Navdeep Jaitly, Quoc Le, Oriol Vinyals, Ilya Sutskever, David Sussillo, Samy Bengio
論文地址:http://papers.nips.cc/paper/6594-an-online-sequence-to-sequence-model-using-partial-conditioning.pdf
12.論文:Attend, Infer, Repeat: Fast Scene Understanding with Generative Models
作者:S. M. Ali Eslami, Nicolas Heess, Theophane Weber, Yuval Tassa, David Szepesvari, Koray Kavukcuoglu, Geoffrey Hinton
論文地址:http://papers.nips.cc/paper/6230-attend-infer-repeat-fast-scene-understanding-with-generative-models.pdf
13.論文:Bi-Objective Online Matching and Submodular Allocations
作者:Hossein Esfandiari, Nitish Korula, Vahab Mirrokni
論文地址:http://papers.nips.cc/paper/6085-bi-objective-online-matching-and-submodular-allocations.pdf
14.論文:Combinatorial Energy Learning for Image Segmentation
作者:Jeremy Maitin-Shepard, Viren Jain, Michal Januszewski, Peter Li, Pieter Abbeel
論文地址:http://papers.nips.cc/paper/6595-combinatorial-energy-learning-for-image-segmentation.pdf
15.論文:Deep Learning Games
作者:Dale Schuurmans, Martin Zinkevich
論文地址:http://papers.nips.cc/paper/6315-deep-learning-games.pdf
16.論文:DeepMath - Deep Sequence Models for Premise Selection
作者:Geoffrey Irving, Christian Szegedy, Niklas Een, Alexander Alemi, François Chollet, Josef Urban
論文地址:http://papers.nips.cc/paper/6280-deepmath-deep-sequence-models-for-premise-selection.pdf
17.論文:Density Estimation via Discrepancy Based Adaptive Sequential Partition.
作者:Dangna Li, Kun Yang, Wing Wong
論文地址:http://papers.nips.cc/paper/6217-density-estimation-via-discrepancy-based-adaptive-sequential-partition.pdf
18.論文:Domain Separation Networks
作者:Konstantinos Bousmalis, George Trigeorgis, Nathan Silberman, Dilip Krishnan, Dumitru Erhan
論文地址:http://papers.nips.cc/paper/6254-domain-separation-networks.pdf
19.論文:Fast Distributed Submodular Cover: Public-Private Data Summarization
作者:Baharan Mirzasoleiman, Morteza Zadimoghaddam, Amin Karbasi
http://papers.nips.cc/paper/6540-fast-distributed-submodular-cover-public-private-data-summarization.pdf
20.論文:Satisfying Real-world Goals with Dataset Constraints
作者:Gabriel Goh, Andrew Cotter, Maya Gupta, Michael P Friedlander
論文地址:http://papers.nips.cc/paper/6316-satisfying-real-world-goals-with-dataset-constraints.pdf
21.論文:Can Active Memory Replace Attention?
作者:Łukasz Kaiser, Samy Bengio
論文地址:http://papers.nips.cc/paper/6295-can-active-memory-replace-attention.pdf
22.論文:Fast and Flexible Monotonic Functions with Ensembles of Lattices
作者:Kevin Canini, Andy Cotter, Maya Gupta, Mahdi Fard, Jan Pfeifer
論文地址:http://papers.nips.cc/paper/6377-fast-and-flexible-monotonic-functions-with-ensembles-of-lattices.pdf
23.論文:Launch and Iterate: Reducing Prediction Churn
作者:Quentin Cormier, Mahdi Fard, Kevin Canini, Maya Gupta
論文地址:http://papers.nips.cc/paper/6053-launch-and-iterate-reducing-prediction-churn.pdf
24.論文:On Mixtures of Markov Chains
作者:Rishi Gupta, Ravi Kumar, Sergei Vassilvitskii
論文地址:http://papers.nips.cc/paper/6078-on-mixtures-of-markov-chains.pdf
25.論文:Orthogonal Random Features
作者:Felix Xinnan Yu, Ananda Theertha Suresh, Krzysztof Choromanski, Dan Holtmann-Rice, Sanjiv Kumar
論文地址:http://papers.nips.cc/paper/6246-orthogonal-random-features.pdf
26.論文:Perspective Transformer Nets: Learning Single-View 3D Object Reconstruction without 3D Supervision
作者:Xinchen Yan, Jimei Yang, Ersin Yumer, Yijie Guo, Honglak Lee
27.論文:Structured Prediction Theory Based on Factor Graph Complexity
作者:Corinna Cortes, Vitaly Kuznetsov, Mehryar Mohri, Scott Yang
論文地址:http://papers.nips.cc/paper/6485-structured-prediction-theory-based-on-factor-graph-complexity.pdf
28.論文:Toward Deeper Understanding of Neural Networks: The Power of Initialization and a Dual View on Expressivity
作者:Amit Daniely, Roy Frostig, Yoram Singer
論文地址:http://papers.nips.cc/paper/6427-toward-deeper-understanding-of-neural-networks-the-power-of-initialization-and-a-dual-view-on-expressivity.pdf
Demonstrations
標題:Interactive musical improvisation with Magenta
作者:Adam Roberts, Sageev Oore, Curtis Hawthorne, Douglas Eck
我們結合了基於LSTM的迴圈神經網路和Deep Q-learning建立了實時生成音樂序列。LSTM的任務是學習音樂評分(編碼為MIDI,而不是音訊檔案)的一般結構。Deep Q-learning用來改進基於獎勵的序列,如期望的型別,組成正確性和預測人類協作者演奏的內容。基於RNN模型的生成與強化學習的結合是一種生成音樂的全新方式。這種方式比單獨使用LSTM更為穩定,生成的音樂更加好聽。該方法有兩個任務:生成對短旋律輸入的響應,以及實時生成對旋律輸入的伴奏,持續對未來輸出進行預測。本方法在TensorFlow中加入了一個全新的MIDI介面產生即興的音樂體驗,讓使用者可以與神經網路實時互動。
標題:Content-based Related Video Recommendation
作者:Joonseok Lee
這是一個相關影片推薦的展示,種子來源於YouTube上隨機的影片,純粹基於影片內容訊號。傳統的推薦系統使用協同過濾(CF) 方法,在有多少使用者在看了種子影片之後觀看特定的候選影片的基礎之上來推薦相關影片。這種方式沒有考慮影片內容但是考慮了使用者行為。在這個展示中,我們關注的是冷啟動問題,其中種子或者候選影片都是新上傳的(或者未被發現的)。對此我們按照一個基於影片內容的相似性學習問題進行建模,並學習了深度影片嵌入經過訓練去預測真實情況的影片關係(由一個CF基於協同手錶的系統鑑定) ,但僅使用視覺內容。它基於任一新影片內容,將其嵌入進一個1024維的表徵中,同時成對影片的相似性在嵌入的空間中僅當做一個點積來計算。我們發現,被學習的影片嵌入超越了簡單的視覺相似性,並能捕捉複雜的語義關係。
更多的 workshops 和 tutorials 可點選網址:https://research.googleblog.com/2016/12/nips-2016-research-at-google.html