Recently, the news of AlphaGo’s successive defeats of human chess players had a huge impact on the world, and a British AI company called DeepMind, which was acquired by Google in 2014, was also known. Google, one of the world’s most advanced AI companies, changed its future strategy from “Mobile First” to “AI First” at the end of last year.
Ng, a former chief scientist at Baidu, was one of the first three members of Google Brain, a deep learning research team at Google. Google Brain began as a research project for Google X. But it was so profitable that it was spun off from Google X and became a separate division of Google’s parent company. Eric Teller, the former head of Google X, revealed that the Google Brain team was making more money than the entire Google X division cost.
At the Geek Park Rebuild 2017 conference, Google Brain senior researcher Xiaobing Liu introduced his team’s research from TensorFlow to the latest deep learning.
A neural network awakened by data
Now, AI can’t get around deep learning, because deep learning was the original driving force of this WAVE of AI, and deep learning is really a deep neural network. Neural networks start out to mimic the human brain, with neurons, with connections between neurons, and the model is learning the strength of those connections.
Put it into a real problem, like image recognition, give us a picture of a cat, we can tell it’s a cat at a glance, but it needs to be recognized by a machine. The machine was presented with an image, and instead of a cat, it saw a bunch of pixels, each 0.255 dots. For example, if the cat picture is 128×128, the machine will see 128×128, a lot of integer values.
In image recognition, there is a concept called deep convolutional network. The basic meaning is to input an image into the network. There are many layers in the network, and each layer will learn different information in the image, such as the first layer is color, the second layer is shape, and the third layer is pattern…… In short, a deep network is a deep layered structure, and after a lot of learning, the machine can figure out whether this is a cat or a dog.
That’s deep convolutional networks.
Neural networks are not a new concept. They were invented in the 1980s and 1990s, but AI didn’t catch on then. Because there wasn’t enough data, and the hardware wasn’t powerful enough. After 40 years of Moore’s Law, CPUS and Gpus are becoming more and more powerful. In the era of big data, we have collected more and more data. At this time, the effect of neural network is much better than other algorithms.
We can see speech recognition, image processing, including unmanned driving, AlphaGo and other things, all thanks to the leap or breakthrough of neural networks.
A strong infrastructure can drive new research and applications
At the end of last year, Google changed its future strategy from “Mobile First” to “AI First,” turning the whole company to AI.
TensorFlow was originally developed by the Google Brain team for machine learning for a variety of perceptual and language understanding tasks. Voice recognition, Gmail, Google Albums, and search — anything that involves intelligence, Google is using deep learning and machine learning to further improve the user experience and product accuracy. TensorFlow was released in 2015 as an open source software. It is widely used in a variety of fields and has entered the classrooms of many universities, such as Stanford, Berkeley, and some top research institutions: OpenAI, DeepMind, and many companies in China are using Tensorflow.
In June 2016, Jeff Dean, head of Google Brain, said there were 1,500 libraries on GitHub that mentioned TensorFlow, of which only five were from Google.
In addition to improving the experience of existing products, Google is also trying something new. AlphaGo, which beat the top human player twice last year and again this year, has benefited from the development of deep learning.
Of course, there is enough hardware to support it. At this year’s Developer conference in May, Google released the second generation OF TPU, a Cloud computing hardware and software system, called Cloud TPU, mainly to improve machine learning computing processing requires a lot of load, including training and reasoning, for Google Cloud computing platform brought a great boon.
Thanks to Google’s strength in hardware and software integration, TensorFlow has become one of the leading platforms for building AI software. As a dedicated chip designed specifically for machine learning, the first GENERATION OF TPU is used by the AlphaGo AI system as the basis for its prediction and decision making.
Liu xiaobing said, “After 40 years of rapid development, Moore’s Law has been limited by some limitations of physics. We need to apply hardware in a more rational way and rethink hardware design based on deep learning itself, which is why we do TPU.”
The bigger the model, the more improvement the neural network will bring
As AI continues to evolve, it will open up new possibilities, such as driverless cars. Because deep learning now allows a computer to see and recognize and understand, in a car, a driverless car. The trend is that there will be more data, models will get bigger and bigger, and neural networks will bring more and more improvements.
After all these years of AI development, the first thing that needs to be solved is the platform problem. Liu Xiaobing said, “Many startups don’t have such resources, but what we do is to let computers generate algorithms automatically, which is called Learn to Learn, or Auto ML. The basic idea is to constantly improve the machine learning effect like AlphaGO. “You could end up with a bunch of data, a bunch of goals, and the machine will do it automatically.”
“For Google depth study, we want to layout is a model of a very large, with automatic generation algorithm, and finally the mapping by other software to the hardware machine learning algorithms to the power of collaborative software and hardware, enable it to achieve the overall effect optimal. We think that to solve some artificial intelligence is a very good solution.”