Yukei: Entering the age of deep learning

Just end of the Geek Park, Baidu in-depth study researcher executive vice President Yukei share the depth of learning technology for Baidu's technology and products brought about by the changes. Today's Future forum, Yukei again to 36 Krypton introduced the recent Baidu in the field of AI and intelligent hardware research progress.

Deep learning: Baidu can do better than Google and Facebook

In the past two years, Baidu has done a lot of things in the field of deep learning. Baidu's brain has made considerable progress, it is not always learning, evolution and evolution, with considerable engineering capacity. I told you two months ago that we built the world's largest deep learning network with Bai parameters. Today, this data is Chi. On this basis we can do a lot of things.

Baidu in search, advertising, image recognition, speech recognition, these areas, have considerable progress.

In this field of advertising, the entire Baidu Phoenix Nest advertising system has entered the era of in-depth learning, the promotion of ad clicks is very significant, the specific number is a trade secret I can not speak, but the percentage of double-digit increase, so this has a great contribution to the company's revenue.

In the past few years, the development of technology has undergone a process from quantitative change to qualitative change. For example, in the 2010, the world's best results are not ideal, with the development of in-depth learning, each year at a rapid pace of evolution.

At the end of 2014 Baidu did 5.9% error rate, which is the world's highest image recognition accuracy rate. Last year we found that Baidu is the best and Google is second in a single model experiment on object detection.

In the voice of this piece, Baidu voice search in two years time recognition rate increased a lot. Now 10% of search demand in mobile search products comes from voice search, which is growing fast.

We also launched the "Baidu translation." You can speak Chinese in New York Street, immediately speech recognition translated into Chinese text, and finally automatically translated into English. can also take pictures automatically-for example, foreigners come to Shanghai to see Chinese menu, photo recognition into English.

In face recognition technology, the Facebook Artificial Intelligence Laboratory announced last year that it had the best results in the world, with an accuracy rate of 97.35%. A few days ago, my colleagues in Baidu made a more exciting progress, the accuracy of face recognition reached 99.62%.

Researchers at Stanford University recently made a test, and people reached 99.2% on the data. Of course, it must be clarified that in a data pole, the machine's algorithm to achieve human level, but the overall machine than people still have a very big gap.

Baidu car: We want to do it with better algorithms and affordable parts.

Unlike Google's unmanned driving, we recognize the integration of people and cars in automatic driving. This process can be divided into several stages. The first stage is assisted security; The second phase is active security; The third stage is automatic driving in a particular scenario, such as repetitive driving in a large logistics center, an adaptive cruise on the freeway, and then a high degree of automation.

Our costs are not as high as Google's and their sensors are expensive. We want to do it with better algorithms and affordable parts. We did it in the car. Simulation system, using depth learning to develop image recognition technology, with low cost to complete this task.

On the concrete landing, we first have to have the ability to build high-precision maps, which is actually a basic construction. We have to do 10-20 centimeters of scale accuracy, such data will be very valuable. If you can do this, you can provide more accurate navigation. Predictably, in the next 1-2 years, the entire navigation system will be changed dramatically. For example, when the overpass, sometimes facing 3-5 of the front fork, look at two-dimensional maps difficult to find the right direction. And we're going to get to the driveway level so we can navigate more accurately.

Baidu Eye: This year there will be exciting progress

The iteration of the Baidu eye is very fast, and it has made great progress compared to the last introduction to Krypton 36. It's more like following your robot, not just positioning, but also knowing what the wearer is looking at, and expanding your information space on this basis. For example, in a museum scene, when you see an oil painting, it can tell you about the painting's author, style, historical background and other information. Computer to mobile phone is a leap, from mobile phone to glasses will be another very important leap. Baidu is doing some advanced thinking and trying, this year Baidu Eye will have exciting progress. Of course, it will still be in the vertical field.