English Original:The Google Brain team & mdash; Looking Back on 2016
Google Brain TeamThe long-term goal is to create more intelligent software systems to improve human life, and through a variety of different areas of pure applied research to achieve. While this is clearly a long-term goal, we would like to step back, review some of the team's progress last year, and share our expectations for 2017.
Research results published
An important way to evaluate the quality of research is through a conference in the top international machine learningICML, & Lt;NIPSwithICLROn the published results. Over the past year, our team has received 27 papers at these conferences, covering a wide range of topics, includingProgrammed, & Lt;Knowledge transfer between networks, & Lt;Distributed training of machine learning model, & Lt;Language model generation, & Lt;Unsupervised Machine Learning, & Lt;Proof of automation theorem, & Lt;Better understanding of neural network theory, & Lt;Improved algorithm for reinforcement learningand many more. There are more papers accepted in other areas of the meeting, such as natural language processingACLwithCoNNLConference, voiceICASSP,VisualCVPR, The robot'sISERAnd computer systemsOSDIGeneral Assembly. Our team also presented 34 papers to the forthcoming ICLR 2017, a top-level conference on in-depth study. able to passHereLearn about our work in these papers.
Natural Language Understanding
One of the key areas of our research is to make computers better understand human languages. At the end of 2014, three researchers from the brain team published a series on sequence-to-sequence neural network learning (Sequence to Sequence Learning with Neural Networks), The feasibility of machine translation technology is demonstrated. In 2015, we show that this technology can also be usedGenerates a picture title, & Lt;Statement analysis,as well asSolving computational geometry problems. By 2016, this prospective study (with a greater advancement) has never allowed brain team members to work closely with the Google translation team, and ultimately with a complete end-to-end learning systemResearch Papers) AllInstead of the Google translation translation algorithm developed. This new system makes up for the gap between the old system and the quality of human translation, with some 85% accuracy for translations between languages. A few weeks later, we show how the system is "Zero-shot tranaslate", That is, without the translation of the sample scenario translation of the study (Research Papers). The system is now deployed in the Google translation environment and supports more and more languages between the translation, to provide users with better quality of translation, so that people eliminate the language barrier more efficient communication. Written by Gideon Lewis-Kraus & ldquo;The Great A.I. Awakening"This in-depth translation of the masterpiece (and the depth of learning and the history of Google's brain team), this in-depth article published in December 2016 in the New York Times.
The commercial robot control algorithm is very carefully and carefully hand-programmed, and because it means bringing new capabilities to the robot, it is usually a very laborious process. We believe that a better solution is for robots to learn and acquire new skills automatically through machine learning. Last year, we worked withGoogle X TeamOf the researchers to show how the robotic arm passesLearning hand-eye coordination, Summing up their experience to self-study faster (Research Papers). Our robots completed about 800,000 well-tried attempts during the study. At the end of the year, we explored three possible scenarios to learn new skills: reinforcement learning, interactive learning, and demonstration learning. We will continue to work toward this goal, allowing robots to learn new tasks in a flexible and easy way, and to run in complex, real-world environments. To help other robotics researchers, weMultiple robot data sets are open.
It is very exciting to be able to enhance the ability of doctors and medical practitioners through machine learning. Just as an example of the possibility, published in the Journal of the American Medical AssociationJAMA) On the onepaper, Showed a machine learning-driven diagnostic system through a retinal image of diabetic retinopathy diagnosis, its ability has been with a certified ophthalmologist. More than 400 million people may be at risk of blindness if they can not detect early symptoms of diabetic retinopathy, but in many countries there are only a handful of ophthalmologists who are screened, and this technique can help ensure that more people are screened. We are doing some work in other areas of medical imaging, as well as in researching machine learning for many more medical forecasting tasks. we believeMachine learning can improve the quality and efficiency of medical experience, 2017 in this area we will have more work to progress can be introduced.
Music and artistic creation
Technology tends to help people create and share multimedia - mdash; for example, print publishing, film, or electro-acoustic guitar. Last year we started a callMagentaOf the project, toDigging the possibilities between art and machine intelligence, And the possibility of using machine learning systems to stimulate human creativity. From the beginning of music and picture generation, to areas such as text generation and VR, Magenta has a state-of-the-art level in the creation of content creation models. We helped organize a topic on this topicOne day seminar, andSupporting a machine-generated art exhibition. inMusic generationwithArt style conversion, To explore a large number of topics,Our jam session demo won the NIPS 2016 Best Presentation Award.
Security and Justice
When we develop more powerful and complex AI systems and apply them to all aspects of the real world, we want to make sure that these systems are safe and just, and we want to build tools to help humans better understand the output they produce. In the area of AI security, the interagency collaboration involving Stanford, Berkeley and OpenAI, we publishedA white paper on AI security specific issues(See alsoHere's the blog post). The paper outlines some specific issues and areas, and believes that there are some really grounded research that will be addressed in the AI security arena. We have made progress in one of the areas of security, namely privacy protection in training data, access toDifferentiated privacy guarantees, Most of which were recently adoptedKnowledge transfer technologyobtain. In addition to security, when we began to trust the AI system to make more complex and accurate decision-making, we want to confirm that these decisions are fair. inA paper on equal opportunity in supervised learning(See alsoHere's the blog post), We show how to optimally adjust any trained predictor to avoid bias and discrimination. This is illustrated by a case study based on the FICO credit score. In order to make this work easier to understand, we also createdA visualization product that aids in explaining and interactively explores the ideas in the paper.
November 2015 usOpen source TensorFlow the initial version, So that other colleagues in the machine learning community can benefit from it, and work together to improve. 2016 TensorFlowBecoming the most popular machine learning program on Github, More than 570 developers have submitted more than 10,000 comments.TensorFlow Knowledge Base ModelBenefit from community contributions have also grown, now only on GithubMore than 5,000 TensorFlow-related code repositories. In addition, TensorFlow has beenWell-known research institutions and large enterprisesincludeDeepMindWidely adopted and applied, and even some special application scenarios, such asSearch in Australia and New Zealand also cattle