About AI innovation and entrepreneurship - the road of "double innovation" in artificial intelligence | AI Technology Review Weekly

Latest update time：2016-10-30

Reads：

AI Technology Reviewer is still not satisfied with the CNCC 2016 conference which just ended last weekend. Whether it was researcher Bao Yungang's proposal to reduce the cost of cloud computing data centers through labeled von Neumann structure, or researcher Shan Shiguang's introduction to the application of deep learning in deep face detection and recognition technology, both demonstrated the academic innovation achievements in the field of artificial intelligence in China.

Li Zhigang, the founder of Mobvoi, who is from the industry, talked about the pitfalls and opportunities on the road to AI entrepreneurship. As one of the earliest entrepreneurs in the field of artificial intelligence, he brought different thoughts and insights to the readers.

In addition to sharing about innovation and entrepreneurship, AI Technology Review also noticed the academic developments of technology giant Google. Its latest enhanced style transfer algorithm can simply allow a single deep convolutional style transfer network to learn multiple artistic styles at the same time.

After talking about so many academic and industry dynamics about artificial intelligence, you are still an AI "novice" and don't even understand what machine learning is. What should you do? Don't worry, AI Technology Review has prepared a benefit for you this week - 16 must-see machine learning video tutorials for novices.

Let’s review this week’s AI technology review headlines.

● ● ●

Analysis of Bao Yungang's 37-page PPT from the Chinese Academy of Sciences: How to reduce the cost of cloud computing data centers?

Bao Yungang is a researcher, doctoral supervisor, and deputy director of the Advanced Computer Systems Research Center at the Institute of Computing Technology, Chinese Academy of Sciences. As a special guest of CNCC 2016, he proposed reducing the cost of cloud computing data centers by labeling the von Neumann structure.

The goal of labeling the von Neumann structure is actually:

To improve the resource utilization of the data center and reduce the overall cost of the data center.

The three key words surrounding this goal are cloud computing, von Neumann structure, and labeling.

But some people also asked me, if it really becomes mainstream in the future, what will the equipment vendors do if the cost of data centers is reduced? Why are they motivated to do it? Because their income seems to be declining.

I would like to mention that 200 years ago, British economist Jevons asked this question: if coal utilization doubled, would coal production decrease? In fact, he found that coal production increased because more people would use it, so technological progress would increase technology consumption. I believe that through our efforts to reduce the cost of cloud computing, we will further promote the rapid development of cloud computing.

● ● ●

Shan Shiguang from the Chinese Academy of Sciences: Deep face detection and recognition technology - progress and prospects

Shan Shiguang is a researcher at the Institute of Computing Technology of the Chinese Academy of Sciences and executive deputy director of the Key Laboratory of Intelligent Information Processing of the Chinese Academy of Sciences. He delivered a speech at the CNCC 2016 Visual Media Computing Forum, introducing the application of deep learning in several key processes in the field of face recognition in recent years.

Special mention should be made of the academic milestones (databases): ORL - FERET - FRGC v2.0 - LFW

The current application status of face recognition is also given:

The report can be summarized as follows:

Face detection and recognition are no longer special
Deep models (& big data & high performance computing) have greatly advanced face recognition capabilities, and can outperform the human eye in some (user-coordinated) tasks
Blacklist video surveillance scenarios involving tens of thousands of people are not yet mature
SeetaFace provides a good baseline for everyone

● ● ●

Mobvoi Li Zhifei: Pitfalls and opportunities on the road to AI entrepreneurship

In a special report at the CNCC 2016 conference, Li Zhifei, founder of Mobvoi, introduced the two most common AI industrialization routes:

Implementing AI-first strategy in existing products, such as Google;
AI is provided to third parties as a technical API, such as Mobvoi.

“The former is suitable for large companies, and the latter is suitable for small companies.”

The advantages of To B are: you can make money in the early stages; you are more focused and don’t need a full-stack team; and you can achieve richer applications.

The disadvantage of To B is that it is difficult to expand in scale.

The advantages of To C are: building an independent brand; gradually forming scale and business model.

The disadvantages of To C are: it requires a longer period of accumulation and a full-stack team.

The price of choosing To C is to go through the painful process from soft to hard. Li Zhifei shared with everyone several aspects that need to be paid attention to:

The laws of the hardware must be discovered and respected;
Must follow the hardware life cycle;
Personalized interactive innovation in hardware is difficult to achieve.

● ● ●

Google's latest enhanced style transfer algorithm

Google released this week a new transfer network (from its paper "A Learned Representation for Artistic Style") that is a simple way to learn multiple styles at the same time, which can simply allow a single deep convolutional style transfer network to learn multiple artistic styles at the same time.

This method can achieve real-time style interpolation, which can be applied not only to static images but also to videos.

As shown in the figure above, in actual use, users can use 13 different painting styles and adjust the relative strength of these styles through sliders. Multiple styles are combined together in real time to get one output.

The following picture shows the result of combining 4 styles in different proportions:

● ● ●

16 Must-see Machine Learning Video Tutorials for Beginners

Many of us don’t realize that there are actually a lot of free machine learning tutorials on YouTube. You don’t have to wait for MOOC courses to be updated. You can find what you want on YouTube. (Note: Please bring your own ladder to watch the Internet scientifically)

This article can help you discover new tools, techniques, methods, etc. You must remember this sentence: the need for learning new knowledge should be as urgent as the need for living water, and never stop chasing new knowledge and new ideas.

These videos are mainly divided into 4 parts, content directory:

1. Getting Started with Machine Learning