Technology: Google's artificial intelligence

 

Technology: Google's artificial intelligence deep dissection

Google's artificial intelligence technology

The era of artificial intelligence technology has arrived - Google is based on AI, integrates it into life, and makes the impossible possible

As early as HAL9000 in Stanley Kubrick's "2001: A Space Odyssey" in 1968, to R2-D2 in "Star Wars" in 1977, to David in "AI" in 2001, to the recent "Star Wars: The Force Awakens BB-8, the countless movie robots, rely on the forward-looking creations of Hollywood visionaries to bring us closer to artificial intelligence.

From AlphaGo and Lee Sedol at the game of Go to various smart products, including Google Home, Google Assistant, and cloud computing hardware, Google has officially established a corporate strategy that prioritizes artificial intelligence. AI business covers everything from hardware to software, search algorithms, translation, speech and image recognition, unmanned vehicle technology, and medical and pharmaceutical research. 

The ultimate goal of AI is to imitate the operation of the brain, and the GPU promotes the popularization of AI, but three major problems still need to be solved

The ultimate goal of artificial intelligence is to imitate the thinking and operation of the human brain, but now the more mature supervised learning (Supervised Learning) does not follow this model. But in the end, unsupervised learning is the most natural way for the human brain to learn.

We believe that in the past 5-10 years, artificial intelligence has been commercialized and popularized mainly due to the rapid increase in computing power: 1) the breakthrough of Moore's Law, which has accelerated the decline in hardware prices; 2) the popularity of cloud computing, And 3) the use of GPUs has improved multi-dimensional computing capabilities, which has greatly promoted the commercialization of AI.

There are three major problems in machine learning:

1. Need to rely on a large amount of data and samples to train and learn;
2. Learning in a specific sector and domain (domain and context-specific);
3. It is necessary to manually select the data expression method and learning algorithm to achieve optimal learning.

The market value of Google is seriously underestimated, and the rise of the lunar exploration business will usher in a new golden decade

The core of the company's profitability is the artificial intelligence-driven search and advertising business. Although the advertising business still accounts for 90% of revenue, with the rise of the Other Bets business in 3-5 years, Google will usher in a new golden decade. Now the market has been comparing Facebook and Google. Google's 2017 PE was 19x, compared to FB's 22x, we think Google is seriously undervalued. In Google's advertising business, the increase in the proportion of mobile compared to the decrease in the proportion of PC is a transition period for the new normal. The huge growth potential of 2B cloud computing and YouTube and the development of artificial intelligence are also far ahead of FB. The high-speed revenue growth of the lunar exploration business also proves that Google's innovation ability has continued to increase. Relying on the accumulation of artificial intelligence and the lunar exploration business will rise one by one in 3-5 years, Google can be regarded as a VC investment portfolio for a long time, even if only one or two projects are successful, the market value can be greatly increased in the future. We think 23x PE in 2017 is reasonable, with a target price of $920 and a "buy" rating.

The Soul and Backbone of Google: Artificial Intelligence Technology

Google's artificial intelligence business covers everything from hardware to software, search algorithms, speech and image recognition, translation, unmanned vehicle technology to medical and drug research, and is the soul and backbone of the company. Here we sort out the various businesses of Google AI.

1.    Google Brain Neural Network Project

The project Google Brain, created in 2011, is the cradle of many familiar projects, including TensorFlow, Word Embeddings, Smart Reply, Deepdream, Inception, and Sequence-to-sequence. Deep learning is critical to Google and is a very efficient tool for training computers to recognize images and videos, giving them human-like capabilities in face judgment, object recognition, and natural language analysis.

Google Brain Neural Network Project

2.    The second-generation machine learning open-source platform: TensorFlow

In November 2015, Google announced to open source TensorFlow under the Apache 2.0 open source license. The operating principle of TensorFlow is to transmit the data represented by the structure tensor (Tensor) to the artificial intelligence neural network for analysis and processing. Its performance is up to 5 times faster than the first generation artificial intelligence system DistBelief. In April 2016, Google's DeepMind, which developed the AlphaGo Go robot, announced that all future research projects would use the TensorFlow platform.

TensorFlow

3.    Google's latest search algorithm: RankBrain

RankBrain, an automated artificial intelligence search system that was added to Google's search algorithm in October 2015, has become the third most important part of the entire algorithm in just a few months. The RankBrain system helps Google process search results and provide relevant information, and can handle 15% of the previously unhandled search requests in Google searches every day.

Google RankBrain

4.    Google Driverless Cars and Google Drivers

At the end of 2014, Google proposed the concept of a prototype of an unmanned vehicle without a steering wheel and brakes. It is designed to be fully unmanned. The main components include a set of LIDAR (Light Detection and Ranging) sensor computing systems composed of 64 laser units. The driverless car is powered by a navigation and map-scanning system that costs between $75,000 and $85,000. The unmanned vehicle is equipped with the Google Chauffeur artificial intelligence control system. When the camera and the LIDAR sensing system scan the surrounding environment of the vehicle and input it into the computer, the computer system judges the type of the object according to its shape, size, movement form, and other characteristics.

Google Driverless Cars

5.    The Combination of Machine Learning and Machine Vision: Image Recognition

In 2013, Google announced that it would add computer vision or machine vision (Computer Vision or Machine Vision) and machine learning technology to its image search function. Users only need to enter the name of the query item to get the corresponding photo search results. The image recognition technology is based on the architecture of the first-generation deep learning system DistBelief. The core technology is the redesigned convolutional neural network and distributed learning. Compared with the neural networks of other teams, the neural network architecture reduces the parameter settings by more than 10 times.

Image Recognition

6.    Natural Language Understanding Open Source Platform: SyntaxNet

In May 2016, Google open-sourced its natural language understanding (NLU) platform SyntaxNet based on machine learning platform TensorFlow and released Parsey McParseface, a training parser (English Parser) program for English. SyntaxNet uses a deep neural network to solve the problem of language ambiguity. After inputting the sentence to be analyzed, SyntaxNet processes each word from left to right, and gradually adds the dependencies between the analyzed words.

7.    Natural Sentence Understanding and Machine Translation: Gmail/Inbox Smart Reply

Gmail/Inbox Smart Reply, launched in November 2015, uses deep learning technology to write email replies. Gmail uses machine learning technology to identify emails that require a user's response and provides three suitable candidate response answers.

8.    The AI behind Allo’s Smart Reply

In Allo, apps can also generate smart reply options from the user's conversation transcript. The Allo team used a method similar to the "encode-decode" two-step model, first using a recurrent neural network to encode the dialogue sentence word by word to generate the corresponding password (token). Then the password enters a long-short term memory (LSTM) neural network to generate a continuous vector, which is further passed through the softmax model to generate a discrete semantic class (discretized semantic class). The next step for the Allo team is to use a second recurrent neural network to pick the most appropriate response from the set of selectable words.

9.    Google Translate: Machine Translation Systems and Image Recognition

Google released Google Translate 10 years ago, and the core algorithm behind it is Phrase-Based Machine Translation (PBMT). The neural machine translation system (NMT) used by Google this time treats the entire sentence as the basic input unit for translation.

Google Translate

10. DeepMind's Deep Q-Network (DQN): imitating the experience playback of the human hippocampus

DeepMind published a paper "Deep Reinforcement Learning at the Human Control Level" in Nature in February 2015, describing the deep neural network Deep Q-Network (DQN) is developed to transform Deep Neural Networks ) combined with reinforcement learning (Reinforcement Learning) deep reinforcement learning system (Deep Reinforcement Learning System). Q-Network is a model-free reinforcement learning method, which is often used to make optimal action selection decisions for finite Markov decision processes.

11. DeepMind launches text-to-speech system WaveNet

DeepMind also launched WaveNet, new research in the field of computer speech synthesis. This is a text-to-speech (TTS) system that uses a neural network system to model the raw audio waveform (Raw SoundWave). DeepMind says the audio quality generated by WaveNet reduces the gap between computer output audio and natural human speech by 50%, surpassing all previous text-to-speech systems. 

12. DeepMind's medical exploration using image recognition technology

More than 3000 ophthalmic optical coherence tomography (OCT) scans are performed every week at Moorfields Eye Hospital in London, England. The hospital provides DeepMind with 1 million eye scan image data of patients, as well as routine diagnosis and treatment measures. 

13. Large-scale machine learning for drug discovery

In February 2015, Google and Stanford University jointly submitted a paper discussing "Large-scale Multitasking Networks for Drug Discovery". Google is collaborating with a Stanford University lab to explore how to use data from multiple sources to improve accuracy in choosing which compounds are effective in treating diseases. Going a step further, Google measured different amounts and types of biological data from multiple disease treatments to improve the predictive accuracy of virtual drug screening. 

14.YouTube video thumbnails with counting machine vision

In October 2015, Google announced the launch of a computer vision technology that uses deep neural networks in image and video classification and recognition to bring better abbreviations to YouTube. When choosing a video thumbnail, the program's "quality model" scores the images, and Google now adds deep neural network technology to this model.

15. The ultimate solution for machine learning computing power: quantum computing

In May 2013, Google announced the official establishment of its Quantum Artificial Intelligence Laboratory jointly established with NASA's Ames Research Center and the Universities Space Research Association (USRA). The lab, located at NASA's Moffett Federal Airfield in Silicon Valley, California, houses a D-Wave 2 quantum computer that Google bought from quantum computer maker D-Wave Systems. Google's goal is to use the powerful computing power of quantum computers to fully explore technologies in the fields of artificial intelligence and machine learning and build better learning models for research in weather forecasting, disease treatment, search algorithm improvement, and speech recognition. Quantum computers can solve such problems more efficiently, that is, skip the local optimal solution and directly find the optimal solution. 

16. Self-developed AI hardware: Tensor Processing Unit TPU

At present, Google's cloud platform already has cloud machine learning, computer vision API, and language translation API, so that all users who use Google's cloud computing platform can use the machine learning system that Google has been using.

To support the operation of its cloud platform, Google has designed a hardware device customized for artificial intelligence operations, the Tensor Processing Unit (TPU) chip. The chip is an integrated chip tailored for machine learning and TensorFlow built on an ASIC chip.