Compile | AI technology base (rgznai100)
Participate in | pigeons, Shawn
Guide language: Last night, Google Brain hosted its annual online Q&A session on Reddit, in which Jeff Dean, Google Brain’s chief scientist, Vincent Vanhoucke, technical lead of the Google Brain team, and his colleagues enthusiastically answered questions. Here are some representative questions to learn about how to apply to Google’s internship program, what to do, application and interview tips, what interns do on a daily basis, and what the brains of Google think about the future of fields like deep learning.
About internship and career choice:
Q: A few months ago I saw an application for your internship program on your website. It said that the program does not require applicants to have a strong machine learning background. As a biostatistical epidemiologist, my main work is health survey, and I really want to apply for this project. My question is: How many of the applicants who are ultimately accepted do not come from machine learning backgrounds, and how does their training differ from that of machine learning researchers?
Jeff Dean(Google Brain) : Of the 27 interns in the first internship program, about a third had computer science backgrounds, a third had math, statistics, or applied math backgrounds, and a third had strong backgrounds in STEM fields like neuroscience, computational biology, and so on. It’s a similar story for the 35 interns accepted this year, and in fact one of them is a PhD in epidemiology. Almost all of the interns had experience with machine learning, although they had never received academic training on ML.
Q: I have a few questions about the Google Brain internship program. The first batch of interns came from a variety of backgrounds: recent graduates, experienced software engineers, and doctoral students
-
Does this mean there is an admission quota for each background? Experienced software engineers, for example, do not compete with PhD students for internships.
-
What qualities do you look for in applicants from different backgrounds? Among these backgrounds, I am particularly interested in recent graduates and experienced software engineer backgrounds.
-
When will the next batch of applications be opened? It’s already September.
Sallyjesm (Google Brain) : We don’t select a certain number of interns based on background or experience level. The backgrounds of interns this year are so diverse that we haven’t set any specific quotas. What we look for is whether the applicant has great research potential and can grow greatly in such a project. We look for technical ability and research interests in applicants, not specific qualifications. The next batch of applications will open on October 2 and close in early January. Please submit the complete application and provide links to your previous work on GitHub or other previous work in the ML field.
Q: I have just started graduate school and have taken ML courses. I hope to participate in the Google Brain internship program. In the future, I hope to enter the field of ML/AI research (focusing on fundamentals and theories, with few applications). I have worked as an intern software engineer in 4 large companies and made some applications on ML. What can I do to increase my chances of working on Google’s Brain team over the next year or two?
Sallyjesm (Google Brain) : Congratulations on becoming a graduate student! Based on my discussions with some interns, I think it’s important to get your hands on projects during graduate school. Because of this experience, interviews can be a lot of fun, both for you and the Google Brain interviewer.
From a practical point of view, you have complete control over two issues:
-
Prepare complete application materials;
-
Complete the application. Be sure to provide the hiring team with all required materials (for example, to avoid failure to submit required documents on time). If your application is unsuccessful in the first year, please consider applying again after gaining more experience.
Q: What’s it like working on the Google Brain team? What’s your daily job? What are your reasons for deciding whether someone is a good fit for your team?
Sara_brain (Google Brain) : I’m an intern at Google Brain. There were 35 interns recruited this year, and we all worked in the same area in Mountain View (though some interns worked in San Francisco). At breakfast each day, I usually discuss their research project with another intern. Next, I will read various papers related to my research area (transparency of convolutional neural networks), use TensorFlow coding, and have meetings with my project mentors and partners. Google Brain researchers are very collaborative, so I usually have lunch or dinner with researchers working on similar issues.
Google Brain interns also do some interesting things on a regular basis: visiting scholars and conducting research discussions (often uncovering previously unthought of topics, such as applying deep learning to space exploration); Interns meet every two weeks (which helps us keep up with other interns’ research); Learn about the latest advances in TensorFlow and provide direct feedback; Run tests on thousands of Gpus.
Last year’s interns Colin published a good post, describes in detail his internship experience: http://colinraffel.com/blog/my-year-at-brain.html
Q: How did you get into AI and Google?
Alextp (Google Brain) : I became interested in machine learning when I was an undergraduate, and then I got a PhD. I interned at Google as a PhD student and worked there for a few years before moving to Google Brain. Interestingly, I remember the first time I got serious about machine learning was in a number analysis class, when we were discussing polynomial approximation function interpolation and extrapolation; There are many objects that can be expressed as numerical functions. What else can we extrapolate? The question caught my eye. Later that year, I discovered the science of ML and have been fascinated ever since.
The future forecast
Q: What do you see as the next biggest challenge in ML?
Jeff Dean(Google Brain) : Right now we tend to build machine learning systems that can only do one or a few specific tasks (sometimes very difficult tasks, such as translating one language into another). I think we really need to design a simple machine learning system that can solve thousands or millions of tasks and learn from the experience of doing those tasks to automatically solve new tasks; Activate different modules on the model based on specific tasks. Many problems need to be solved to implement such a system. I gave a talk at the Scaled ML conference at Stanford university earlier this year, and I provide some material on this idea in the slides after page 80 (some background is provided later on page 62).
Vincent Vanhoucke(Google Brain) : Making the deep Web stable for online updates from poorly supervised data is still a big problem. Solving this problem will make true lifelong learning possible and open up many applications. Another big hurdle is that the most exciting developments in fields like GAN or Deep RL have not yet reached the point of “batch normalization” : the point at which everything defaults to “want to train” and no longer needs to be tuned hyperparametric by hyperparametric.
Of course, these advances are not yet mature enough to move from interesting research to technology we can rely on. These models cannot now be trained predictably without a lot of fine tuning, so it is difficult to incorporate them into more elaborate systems. The predictable training of these models makes it difficult to apply them to more complex systems.
Vincent Vanhoucke(Google Brain) : Most robot development in the last 10 years has been based on the premise that robots don’t have any sense. As a result, much of the research in the field has focused on developing robots that work in very restricted environments. Now that we have new computer vision “superpowers,” we can turn the field upside down and rebuild robotic systems centered on sensing the unknown and rich feedback. Deep reinforcement learning, the most likely approach, puts perception at the center of the control feedback loop, but it is still a long way from being widely used.
We need to figure out how to make it easier to assign rewards, more reliable in training, and more efficient in handling samples. I discussed some of these challenges at the AAAI conference. I’m very excited that we can now make the system imitate by learning third party vision to solve the task assignment problem and sample processing efficiency problem. If you’re interested in the field, we’ll be broadcasting the first robotics learning conference in a few months.
Specific technical q&A
Q: Do you plan to support the ONNX (Open Neural Network Switching) interchange format? If there is no such plan, why not?
[1]https://research.fb.com/facebook-and-microsoft-introduce-new-open-ecosystem-for-interchangeable-ai-frameworks/
Jeff Dean(Google Brain) : They posted it on their blog a couple of days ago, and that’s when we learned about it. If this format were useful, I doubt the TensorFlow community would support it. We made our open source announcement in November 2015, and since then the TensorFlow source code base has provided us with a format for preserving and restoring model data and parameters.
Q: Two questions
-
Everyone is talking about the success achieved in ML/Al/DL. Can you talk about some of the frustrations or challenges you’ve had trying to solve (research or real) problems with DL? It is best to address the frustrations and challenges of large supervised learning tasks, where DL approaches are generally feasible.
-
What does the Google Brain team think of current methods of unsupervised learning? Do you think there will be big conceptual advances in the next few years?
Vincent Vanhoucke, Google Brain: Frustration: A few colleagues on our team tried to train a neural network title generator using cartoons from the New Yorker magazine with Bob Mankoff, the magazine’s cartoon editor (I just saw that he had a paper published this year). It didn’t work out very well. The title generated by this generator is not funny at all. Although we don’t have enough data according to DL standards, we can use other types of comics to pre-train visual expression. I still hope we can succeed in this, but maybe we’ll do it the old-fashioned way.
Unsupervised learning: I think people are finally realizing that auto-coding is a terrible idea, that unsupervised learning works and supervised learning doesn’t, and that the difference between the two is that it predicts the causal future (next word or next frame) rather than the present (auto-coding). I was delighted to see so many people benchmarking their “future predictions” with our open source push dataset from last year, which was really unexpected.
Q: Have you tried to create a standard coding approach and/or method for Tensorflow and machine learning? People seem to code models in many different ways, some of which are difficult to explain. This question has little to do with the first question. Keras will join Tensorflow. Will Learn be phased out? It may seem odd to have two different high-level apis for the same library.
Wickesbrain (Google Brain) : My advice is to stick with the highest-level APIS to solve problems. That way, you automatically take advantage of the improvements we make internally, and your code is least likely to become obsolete in the future.
Now that we have the full TF.Keras (in the real sense), we are trying to unify the KerAS application with the previous TF concept. The work will soon be finished. Our goal is that tF.Keras can easily gather all the symbols needed to build a complete Keras API spec from one place. Note that Keras is not suitable for all usage cases, especially distributed training and more complex models, which is why we use TF.estimator.estimator. We will continue to improve Keras’s integration with these tools.
We’ll soon start weeding out some of contrib, including all contrib/learn. A lot of people are still using this tool, it will take time to phase it out, and we don’t want to end it unnecessarily abruptly.
About Learning Resources
Q: How do you keep up with the latest developments in the industry? Specifically, what magazines/conferences/blogs/companies do you recommend that showcase cutting-edge technology?
Jeff Dean(Google Brain) : Paper presented at the Top ML conference; Arix Sanity; The “My Updates” feature on Google Scholar; Interesting papers pointed out and discussed by researchers; Interesting articles discussed on Hacker News or on this subreddit.