Editor's Note: This article is a speech by Google Brain leader Jeff Dean at the Silicon Valley Artificial Intelligence Frontier Forum AI FronTIers.
In the history of deep learning, the neural network method began to exert its effectiveness after 1980-1990. With the help of data volume and computational power, the method of deep learning neural network made us get better research and development. Other methods have higher accuracy (in the fields of image, voice, etc.). Before 2011, the method of deep learning can achieve an image misrecognition rate of 26%, and today the number exceeds the human misunderstanding rate (5%), reaching 3%. The deep learning approach in Google's product line today is very versatile, including the Android platform, various apps, drug research, Gmail, and more.
What is the current outcome of the Google Brain team:
In terms of research, 27 papers were published in various top conferences;
Promote the integration and optimization of Google search, advertising, photo albums, translation, Gmail and other product lines;
Publish open source tools such as TensorFlow that are highly popular in the community.
When it comes to the development of TensorFlow, our initial starting point is to create a suitable deep learning tool.
This tool needs to meet the following conditions:
1. Suitable for machine learning thinking and expression of algorithms;
2. High efficiency and ability to test ideas quickly;
3. Compatibility is good, the experiment can run on different platforms;
4. Share and reproduce research issues in different environments;
5. Suitable for productization: can quickly transition from the research phase to the product application phase;
To sum up, TF's goal is to create a general-purpose system for rapid experimentation in machine learning, and to ensure that the system is both the best for research and productization. Finally, this system is not only Google, but also open source, belonging to everyone on the platform.
On November 9, 2015, we released the initial version of TensorFlow, and we have achieved this result:
1. TF currently has 500+ code contributors;
2. Since its release, there have been 12,000+ code submissions;
3. More than 1 million code base downloads;
4. A large number of schools and businesses have built their research and development work on top of TF (Berkeley, Stanford, OpenAI, Snapchat).
We are constantly updating our support for hardware and software platforms. The data shows that we are already the most popular deep learning tool on GitHub.
What are the important implications of deep learning at Google?
In speech recognition, we have reduced the error rate of word recognition by at least 30%;
Deep convolutional neural networks make it possible to search unmarked photos directly;
We use deep learning methods to capture recognized text in Street View photos;
Also use the deep learning method to retrieve the roof of the solar energy in the satellite overhead view;
In medical imaging, retinal images are used for diagnosis of diabetes;
Machine people can now understand the environment and semantics through machine learning methods;
RankBrian is even used for ranking optimization in Google search;
In Inbox, we automatically recommend possible responses through semantic analysis, and 10% of the current Inbox responses are sent via recommendation generation;
In other aspects of machine learning:
In the past, very good models were obtained from scratch training, which is very inefficient. Our TPU, which is specifically designed for deep learning, will enter mass production in the next 20 months.
In our scenario, the future search request might look like this: Please help me find all the literature on deep learning and robots and summarize them in German.
I think that in the next 3-5 years, through the development of speech recognition and semantic understanding, robots/automobiles will become a very important field in the industry.
PS: With PPT+ text version.
In the history of deep learning, the neural network method began to exert its effectiveness after 1980-1990. With the help of data volume and computational power, the method of deep learning neural network made us get better research and development. Other methods have higher accuracy (in the fields of image, voice, etc.).
Before 2011, the method of deep learning can achieve an image misrecognition rate of 26%, and today the number exceeds the human misunderstanding rate (5%), reaching 3%.
What is the current outcome of the Google Brain team:
In terms of research, 27 papers were published in various top conferences;
Promote the integration and optimization of Google search, advertising, photo albums, translation, Gmail and other product lines;
Publish open source tools such as TensorFlow that are highly popular in the community. When it comes to the development of TensorFlow, our initial starting point is to create a suitable deep learning tool.
This tool needs to meet the following conditions:
Suitable for machine learning thinking and expression of algorithms;
Highly efficient, able to test ideas quickly;
Good compatibility, the experiment can run on different platforms;
Sharing and recurring research issues in different environments;
Sport Earphone
Sports headphones are the perfect companion when running, jogging, hiking, or working out in the gym – with the right music, you'll have twice as much energy and motivation.Sports headphones need to deliver on several counts. They must be resistant to rain and sweat. They should also be tough enough to cope with the movement of exercise, staying firmly and comfortably in place, even during intense workouts.Sometimes,bose soundsport erabuds,bose sport earbuds are used for sport earphone.
Sport Earphone,Sony Sports Earphones,Earphones for Cycling,Sony Sport Earbuds
Pogo Technology International Ltd , https://www.wisesir.net