Google AI Blog: March 2018

Investing in France’s AI Ecosystem

Wednesday, March 28, 2018

Posted by Olivier Bousquet, Principal Engineer, Google Zürichannounced the launch of a new AI research team in our Paris officeDeepMind has also announcedFields MedalistCédric Villanireport on AI

We are sponsoring “Artificial Intelligence and Visual Computing” Chair at École Polytechnique (one of the leading higher education institutions in France) which will support their education initiatives in AI

We just established a partnership with INRIA for conducting collaborative research projects

We are funding academic research with unrestricted grants mostly dedicated to the support of PhD and postdoc positions through our Faculty Research Awards and PhD Fellowship programs, as well as our Focused Research Awards. As one example, we have recently funded a project on large scale optimization of neural networks led by Francis Bach (INRIA and ENS) and Alexandre d’Aspremont (CNRS and ENS)

We are working on offering CIFRE PhD positions (joint PhD positions between Google and an academic lab) as well as internships for PhD students

Cordelia SchmidINRIAVisiting Faculty

Using Machine Learning to Discover Neural Network Optimizers

Wednesday, March 28, 2018

Posted by Irwan Bello, Research Associate, Google Brain TeamDeep learningSearchTranslatePhotosstochastic gradient descentmore advanced optimizersnon-convexGoogle Brain teamAutoMLNeural Optimizer Search with Reinforcement LearningPowerSignAddSignImageNetTensorflowrecurrent neural network10convolutional neural networkCIFAR10reinforcement learning

An overview of Neural Optimizer Search using an iterative process to discover new optimizers.

PowerSignlinear cosine decay

Graph comparing learning rate decay functions for linear cosine decay, stepwise decay and cosine decay.

ConvNetPowerSignAddSignNeural Machine TranslationBLEUNeural Optimizer Searchopen sourcing these optimizers in Tensorflow

Expressive Speech Synthesis with Tacotron

Tuesday, March 27, 2018

Posted by Yuxuan Wang, Research Scientist and RJ Skerry-Ryan, Software Engineer, on behalf of the Machine Perception, Google Brain and TTS Research teamsTacotronprosodyTowards End-to-End Prosody Transfer for Expressive Speech Synthesis with Tacotronprosody embedding

We augment Tacotron with a prosody encoder. The lower half of the image is the original Tacotron sequence-to-sequence model. For technical details, please refer to the paper.

Text: *Is* that Utah travel agency?

Reference prosody (Australian)

Synthesized without prosody embedding (American)

Synthesized with prosody embedding (American)

Reference Text: For the first time in her life she had been danced tired.Synthesized Text: For the last time in his life he had been handily embarrassed.

Reference prosody (American)

Synthesized without prosody embedding (American)

Synthesized with prosody embedding (American)

Text: I've Swallowed a Pollywog.

Reference prosody (Unseen American Speaker)

Synthesized without prosody embedding (British)

Synthesized with prosody embedding (British)

Towards End-to-End Prosody Transfer for Expressive Speech Synthesis with Tacotronthis web pageStyle Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech SynthesisGlobal Style Tokens

Model architecture of Global Style Tokens. The prosody embedding is decomposed into “style tokens” to enable unsupervised style control and transfer. For technical details, please refer to the paper.

Text: United Airlines five six three from Los Angeles to New Orleans has Landed.

Style 1

Style 2

Style 3

Style 4

Style 5

style transferStyle Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesisthis web pageAcknowledgementsThese projects were done jointly between multiple Google teams. Contributors include RJ Skerry-Ryan, Yuxuan Wang, Daisy Stanton, Eric Battenberg, Ying Xiao, Joel Shor, Rif A. Saurous, Yu Zhang, Ron J. Weiss, Rob Clark, Fei Ren and Ye Jia.

Reformulating Chemistry for More Efficient Quantum Computation

Thursday, March 22, 2018

Posted by Ryan Babbush, Senior Research Scientist, Quantum AI TeamAntikythera mechanismclassically intractable

Left: The first known computing device, the Antikythera mechanism: a classical machine used to simulate classical mechanics. Right: Google’s 22 Xmon qubit “foxtail” chip arranged in a bilinear array on a wafer, the predecessor to Google’s new Bristlecone quantum processor with 72 qubits, a quantum machine we intend to use to simulate quantum mechanics, among other applications.

Quantum AI teamfirst quantum chemistry experiment on a superconducting quantum computing devicePhysical Review Xexperimentally simulated exotic phases of matterreleasedOpenFermionBristlecone quantum processorQuantum Information ProcessingIBM ThinkQLow-Depth Quantum Simulation of MaterialsPhysical Review XProfessor Garnet Chan at CaltechQuArC group at Microsoftbasis setsHamiltonianplane wavescrystal of lithium hydridecathodesheterogeneous catalystsgraphenesuperconductorsQuantum Simulation of Electronic Structure with Linear Depth and ConnectivityPhysical Review LettersAspuru-Guzik group at Harvarderror-correctioneighty years

Google Faculty Research Awards 2017

Tuesday, March 20, 2018

Posted by Maggie Johnson, Vice President of Education and University Relations, Google
Google Faculty Research Awardshuman computer interactionmachine learningmachine perceptionsystems

There was a 17% increase in the total number of proposals received

There was a 87% increase in the number of proposals from Asia Pacific universities

Proposals focused on Computational Neuroscience increased 53%

Proposals focused on Quantum Computing more than doubled this round

recipientsour websitehere

Using Deep Learning to Facilitate Scientific Image Analysis

Friday, March 16, 2018

Posted by Samuel Yang, Research Scientist, Google Accelerated Science Teamrobotic microscopy applicationsdistinguish cellular phenotypesseparate signal from noiseDeep Learning to the RescueAssessing Microscope Image Focus Quality with Deep LearningTensorFlowFiji (ImageJ)CellProfiler

A pre-trained TensorFlow model rates focus quality for a montage of microscope image patches of cells in Fiji (ImageJ). Hue and lightness of the borders denote predicted focus quality and prediction uncertainty, respectively.

publicationTensorFlowFijiCellProfilerWhat about Images without Objects?What's Next?datasets and modelsAcknowledgementsWe thank Claire McQuin, Allen Goodman, Anne Carpenter of the Broad Institute and Kevin Eliceiri of the University of Wisconsin at Madison for assistance with CellProfiler and Fiji integration, respectively.

Using Evolutionary AutoML to Discover Neural Network Architectures

Thursday, March 15, 2018

Posted by Esteban Real, Senior Software Engineer, Google Brain Team500 million years agoartificial neural networksfind what's in a photographcall a genetic varianthelp diagnose a diseaseStanley and Miikkulainen 2002OpenAIUber LabsSentient LabsDeepMindGoogle Brain teamAutoMLreinforcement learningevolveLarge-Scale Evolution of Image ClassifiersICML 2017Regularized Evolution for Image Classifier Architecture SearchTPUv2 chipsCIFAR-10ImageNetA Simple ApproachCIFAR-10seedparentchilddies

Progress of an evolution experiment. Each dot represents an individual in the population. The four diagrams show examples of discovered architectures. These correspond to the best individual (rightmost; selected by validation accuracy) and three of its ancestors.

Inception-ResNetsimplearchitectureCombining Evolution and Hand DesignZoph et al.stackcells

The building blocks introduced in Zoph et al. (2017). The diagram on the left is the outer structure of the full neural network, which parses the input data from bottom to top through a stack of repeated cells. The diagram on the right is the inside structure of a cell. The goal is to find a cell that yields an accurate network.

Regularized Evolution for Image Classifier Architecture SearchA Controlled Comparison

Comparison between evolution, reinforcement learning, and random search for the purposes of architecture search. These experiments were done on the CIFAR-10 dataset, under the same conditions as Zoph et al. (2017), where the search space was originally used with reinforcement learning.

regularizationthe paperAmoebaNetsAcknowledgements We would like to thank Alok Aggarwal, Yanping Huang, Andrew Selle, Sherry Moore, Saurabh Saxena, Yutaka Leon Suematsu, Jie Tan, Alex Kurakin, Quoc Le, Barret Zoph, Jon Shlens, Vijay Vasudevan, Vincent Vanhoucke, Megan Kacholia, Jeff Dean, and the rest of the Google Brain team for the collaborations that made this work possible.

Balanced Partitioning and Hierarchical Clustering at Scale

Wednesday, March 14, 2018

Posted by Hossein Bateni, Research Scientist and Kevin Aydin, Software Engineer, NYC Algorithms and Optimization Research Teamgraph partitioningbalanced graph partitioning problemNP-hardbest approximation algorithmssemidefinite programmingwe developedWSDM 2016 paperNIPS 2017 paperBalanced Partitioning via Linear Embeddingaffinity clusteringAffinity Hierarchical ClusteringBorůvka’s classic Maximum-cost Spanning Tree algorithm01201

NIPS’17 paperDHTsThis videoComparison to Previous WorkFENNELSpinnerMETISlabel propagation-based algorithmTwitter followership graphUgander and Backstrom, 2013

UB13
(5%)

Spinner
(5%)

Vanilla Affinity
Clustering
(0%)

Final Algorithm
(0%)

37.0%

38.0%

35.71%

27.50%

43.0%

40.0%

40.83%

33.71%

46.0%

43.0%

43.03%

36.65%

47.5%

44.0%

43.27%

38.65%

100

49.0%

46.0%

45.05%

41.53%

ApplicationsGoogle Mapstreatment groups for experimental designshardsHilbert embeddingAcknowledgementsWe especially thank Vahab Mirrokni whose guidance and technical contribution were instrumental in developing these algorithms and writing this post. We also thank our other co-authors and colleagues for their contributions: Raimondas Kiveris, Soheil Behnezhad, Mahsa Derakhshan, MohammadTaghi Hajiaghayi, Silvio Lattanzi, Aaron Archer and other members of NYC Algorithms and Optimization research team.

Behind the Motion Photos Technology in Pixel 2

Tuesday, March 13, 2018

Posted by Matthias Grundmann, Research Scientist and Jianing Wei, Software Engineer, Google Research
new camera featureMotion Stills for Android

Motion photos on the Pixel 2 in Google Photos. With the camera frozen in place the focus is put directly on the subjects. For more examples, check out this Google Photos album.

Camera Motion Estimation by Combining Hardware and SoftwareMotion StillsFused Video Stabilization

Motion photo as captured (left) and after freezing the camera by combining hardware and software For more comparisons, check out this Google Photos album.

affine transformationhomography

Feature classification into background (green) and foreground (orange) by using the motion metadata from the hardware sensors of the Pixel 2. Notice how the new approach not only labels the skateboarder accurately as foreground but also the half-pipe that is at roughly the same depth.

scene at infinityparallaxmixture homographiesrolling shutter

Background motion estimation in motion photos. By using the motion metadata from Gyro and OIS we are able to accurately classify features from the visual analysis into foreground and background.

Motion Photo Stabilization and Playbacklinear programming techniquesearlier posts

Motion photos stabilize even complex scenes with large foreground motions.

Motion Photo SharingAcknowledgementsMotion photos is a result of a collaboration across several Google Research teams, Google Pixel and Google Photos. We especially want to acknowledge the work of Karthik Raveendran, Suril Shah, Marius Renn, Alex Hong, Radford Juang, Fares Alhassen, Emily Chang, Isaac Reynolds, and Dave Loxton.

Semantic Image Segmentation with DeepLab in TensorFlow

Monday, March 12, 2018

Posted by Liang-Chieh Chen and Yukun Zhu, Software Engineers, Google Researchportrait mode of the Pixel 2 and Pixel 2 XL smartphonesmobile real-time video segmentationimage-level classificationbounding box-level detection

DeepLab-v3+^*TensorFlowconvolutional neural networkPascal VOC 2012Cityscapesdepthwise separable convolution

AcknowledgementsWe would like to thank the support and valuable discussions with Iasonas Kokkinos, Kevin Murphy, Alan L. Yuille (co-authors of DeepLab-v1 and -v2), as well as Mark Sandler, Andrew Howard, Menglong Zhu, Chen Sun, Derek Chow, Andre Araujo, Haozhi Qi, Jifeng Dai, and the Google Mobile Vision team. References

Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation, Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam, arXiv: 1802.02611, 2018.

Xception: Deep Learning with Depthwise Separable Convolutions, François Chollet, Proc. of CVPR, 2017.

Deformable Convolutional Networks — COCO Detection and Segmentation Challenge 2017 Entry, Haozhi Qi, Zheng Zhang, Bin Xiao, Han Hu, Bowen Cheng, Yichen Wei, and Jifeng Dai, ICCV COCO Challenge Workshop, 2017.

Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs, Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille, Proc. of ICLR, 2015.

Deeplab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs, Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille, TPAMI, 2017.

Rethinking Atrous Convolution for Semantic Image Segmentation, Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam, arXiv:1706.05587, 2017.

* DeepLab-v3+ is not used to power Pixel 2's portrait mode or real time video segmentation. These are mentioned in the post as examples of features this type of technology can enable.^↩

Introducing the iNaturalist 2018 Challenge

Friday, March 9, 2018

Posted by Yang Song, Staff Software Engineer and Serge Belongie, Visiting Faculty, Google Researchdeep learningself-driving carsvirtual realityfine-grainedinstance-level recognitioninstance-level landmark recognition challengelong tail2018 iNaturalist ChallengeiNaturalistVisipediaFGVC5CVPR 2018iNat-2017KaggleGitHub repothe following photo

The map on the right shows where the photo was taken. Image credit: Serge Belongie.

Trachemys scriptafine-grainedsubordinatelong-tailed distribution

Distribution of training images per species for iNat-2017 and iNat-2018, plotted on a log-linear scale, illustrating the long-tail behavior typical of fine-grained classification problems. Image Credit: Grant Van Horn and Oisin Mac Aodha.

FGVC5furniture categorization challengeFGVCxFGVC5CVPR 2018competitionsAcknowledgementsWe’d like to thank our colleagues and friends at iNaturalist, Visipedia, and FGVC5 for working together to advance this important area. At Google we would like to thank Hartwig Adam, Weijun Wang, Nathan Frey, Andrew Howard, Alessandro Fin, Yuning Chai, Xiao Zhang, Jack Sim, Yuan Li, Grant Van Horn, Yin Cui, Chen Sun, Yanan Qian, Grace Vesom, Tanya Birch, Celeste Chung, Wendy Kan, and Maggie Demkin.

Open Sourcing the Hunt for Exoplanets

Thursday, March 8, 2018

Posted by Chris Shallue, Senior Software Engineer, Google Brain Team(Crossposted on the Google Open Source Blog)two exoplanetsneural networkKepler space telescopehealthcarequantum chemistryfusion researchrelease our codeK2Transiting Exoplanet Survey Satellite
A Planet Hunting PrimerKepler’s photometer

A light curve from the Kepler space telescope with a “U-shaped” dip that indicates a transiting exoplanet.

binary starstarspotscosmic ray

The first light curve has a “V-shaped” pattern that tells us that a very large object (i.e. another star) passed in front of the star that Kepler was observing. The second light curve contains two places where the brightness decreases, which indicates a binary system with one bright and one dim star: the larger dip is caused by the dimmer star passing in front of the brighter star, and vice versa. The third light curve is one example of the many other non-planet signals where the measured brightness of a star appears to decrease.

Kepler data processing pipelinesignal-to-noise ratiosA Machine Learning ApproachGoogle Brain teamhuman genomessketchesformal mathematical logicAndrew Vanderburg

We trained a convolutional neural network (CNN) to predict the probability that a given Kepler signal is caused by a planet. We chose a CNN because they have been very successful in other problems with spatial and/or temporal structure, like audio generation and image classification.

this paper

Searching for New PlanetsKepler-90 i and Kepler-80 gFind your own Planet(s)!Kepler-90 ihome pageBox Least SquaresVARTOOLSLcToolsPlanet Hunters

A low signal-to-noise detection in the light curve of the Kepler 90 star detected by the BLS algorithm. The detection has period 14.44912 days, duration 2.70408 hours (0.11267 days) beginning 2.2 days after 12:00 on 1/1/2009 (the year the Kepler telescope launched).

python predict.py --kepler_id=11442793 --period=14.44912 --t0=2.2
--duration=0.11267 --kepler_data_dir=$HOME/astronet/kepler
--output_image_file=$HOME/astronet/kepler-90i.png
--model_dir=$HOME/astronet/model
prediction = 0.94our paper

our paperIf you’d like to learn more, Chris is featured on the latest episode of This Week In Machine Learning & AI discussing his work.

Blog