spoken search query, regardless of the accent of the user. While this sounds simple, it is a grand challenge problem in artificial intelligence. Our voice systems, which are available on Android, iPhone and other mobile devices, are trained on over 230 billion spoken utterances and possess a one million word vocabulary—and we are working to make them even better. Interestingly, we have a new challenge: Quite often, our systems are even more accurate than the humans that rate them for accuracy, making it challenging to evaluate our own quality!
In the domain of ultra-large software systems, John Wilkes, a distinguished engineer from Mountain View, spoke to the audience about a new system we are building that will automatically manage the seemingly endless number of computers in Google’s worldwide data centers.
Some of these computers are working round the clock answering user queries, processing e-mail, or otherwise attending to tasks that require instantaneous responses. Other computers are working on long jobs, for example, actually learning to do language translation from vast corpora in English, French, German, Chinese, etc. The key, John described, is to make sure that we can easily specify the requirements of each job that needs to run, and then have an uber-manager, “the cluster management system,” automatically allocate those jobs to the right sets of computers in a way that maximizes performance while minimizing costs. Our current cluster management system is seven years old, and we discussed its success and challenges, as well as our hopes for the new system we are building to replace it.
While you wouldn’t think that professors of computer science would be interested in shopping or commerce, Andrew Moore, a former professor at Carnegie Mellon and now the Director of Google’s Pittsburgh office, described the deep research questions in areas such as shoe shopping. How can we implement a system to analyze the image of a shoe—its color, shape and pattern? How can we show a pair of shoes that someone might purchase based on this image analysis? How can we simultaneously provide accurate results of shoes a shopper would most likely purchase, without showing shoes they would not like, and provide the serendipitous connections that a shopper would experience in a real store? Optimization, computer vision algorithms, auction theory and more all play a role, and are the subject of active research not only at Google, but across the computer science community.
The field of computer science is, in many ways, a large expanding sphere that grows into ever more domains of applicability. The greatest recent successes have been at the boundary of computer science and virtually every other discipline. So, we covered quite a diversity of topics in New York. There is more on our research blog and also on the Official Google Blog, where you can even see a poem by NYU Professor Ken Perlin in iambic pentameter, musing about the future of mobile devices.