Acrofan News

Google AI Forum 6th Round: AI Innovation and Cloud

ACROFAN=Hyung-Keun Kim | Published : Monday, November 6, 2017, 1:21 pm
On the morning of September 12th, Google hosted the 'Google AI Forum 6th Round: AI Innovation and Cloud' at its office in Gangnam-gu, Seoul.

In the upcoming AI-First era, Google AI Forum is an event monthly organized by Google to provide opportunities to a deeper understanding of artificial intelligence and machine learning, with intelligible explanations and examples.

 
▲ Google AI Forum 6th Round: AI Innovation and Cloud was held.

As the first order of the event, Jang Hye-deok, director of Google Cloud Korea, gave an outline of Google Cloud business.

Jang Hye-deok began the presentation by introducing that "Google is a company that works with a mission to gather and organize information around the world and make it accessible to everyone, anywhere, and it starts with a search business, but today, one billion people use services every day, including search, Android, Chrome, Maps, Play, YouTube, and Gmail."

She explained that Google provides a huge scale of Internet services, but it has been collecting and indexing a variety of data and trying to provide more efficiently and understand users through AI. It's easy to see that YouTube is able to receives large videos and provides them to users all over the world after saving them, and Gmail is connected quickly from anywhere in the world and automatically backs up email so users do not lose their email. Likewise, many invisible tasks that provide convenience were possible because of the constant effort Google's engineers and computer scientists have been working on.

Jang Hye-deok said, "As Google has been in business for more than 15 years, it has been through most of the sectors in the computer science field. The main task of Google Cloud is to package it so that external developers and corporate customers can use it depending on their needs."

Four characteristics were mentioned as Google Cloud's strengths. According to the introduction, Google has a global infrastructure and network that connects data centers by installing undersea cables, although it is not a telecommunications company. Google's top-level engineers are in charge of helping customers reduce operational burden and focus more on securing insights. In addition, by the rate system which is tailored to the client's situation, up to 60% cost reduction compared to other cloud can be expected, and it also secures expandability through open source leadership about machine learning.

Lastly, Jang Hye-deok said, "Because of the tight contact points around the globe, customers will be able to get a good user experience through Google Cloud, no matter where they are."

 
▲ Jang Hye-deok, CEO of Google Cloud Korea, introduced the outline of Google's cloud business.

 
▲ There are 7 services that 1 billion people use every day among Google's services.

 
▲ Four features about the strengths of Google cloud are mentioned.

As the second order of the event, Jia Li, director of Google Cloud AI and ML R&D, gave a video lecture on AI innovation and cloud.

Jia Li started the lecture by saying, "Although AI has evolved from academic research, now it is at the center of the greatest change in industry. Many examples show that a number of companies benefit from the efficiency of AI, and due to this effect, AI will be one of the exciting areas."

According to the lecture, especially in the next stage of AI, 'AI democratization' should be done to reach the maximum number of people, and it will lower the entry barriers and give advantage to as many developers, users and companies as possible. Moreover, for everyone to get the most out of it, there is a need to pay attention to key elements such as computing, algorithms, data, talent and expertise.

Google Cloud is taking advantage of computing power's GPU, CPU, and cloud TPU to cover the entire machine running. Among them, the cloud TPU, which was presented at this year's I/O event, is the second generation product of Tensor Processing Unit. The first generation could only be used to run a given machine learning model, while the training and running had to be done on separate hardware. However, the second-generation TPU can both train and run the machine learning model at the same time. Performance is a level of 180 trillion floating point operations in a sec per a single unit. Google is providing cloud TPU to its customers through Google Cloud Engine, enabling research institutes and companies, which utilize machine learning, to obtain new efficiencies and try more things in a short time.

Next, Jia Li said, "Of course, computing power is essential and important, but it is only the first step for utilizing AI. Even if all the computing power available all over the world are secured, AI is still a very complex and challenging area, so companies need to have a variety of tools to use them." And, "Here, the tool may be a machine learning library such as a tensorflow, or it may utilize the model through a pre-trained API."

To provide the pre-trained API, there is a need to organize the data for training, and one of the things Google has prepared for it is ImageNet. ImageNet has more than 150,000 object categories and more than 14 million images. It is recognized as one of the largest visual data sets, and the algorithm built on it has rapidly improved the state of computer vision. Thanks to this, the recognition error rate has drastically decreased, and this improvement not only allows developers to experience through the cloud vision APIs of various services, but also enables developers to utilize their own algorithms.

These trained models are available as tensorflow based services and can be used in large scale machine learning projects. Since this service provides basic infrastructure and secures expandability, for customers who utilize this, it is convenient to focus on only bringing the best results using machine learning model.

Jia Li emphasized the importance of data by explaining, "As human beings learn a lot from their lifetimes, they need to have huge amounts of data in AI to keep them up. Businesses need to learn how to collect, classify, and process meaningful data and implement meaningful projects properly." She also stated that Google shares many types of data sets, including genetic-related public data and YouTube data sets.

Meanwhile, she introduced that Google is making great efforts to educate and invest in talent. Each year, more than 250 research projects from around the world are funded annually by Google, and it is also investing in manpower to provide scholarships for doctoral degrees and to train thousands of interns.

Google's in-house training program encourages Google's engineers to grow their expertise in machine learning. By expanding it to external programs, companies can get AI-related training on Google sites and give them the opportunity to do real work with Google's machine learning specialists.

By finishing the lecture, Jia Li said, "AI is one of the most important technologies of our century, and we will try to keep Google Cloud ahead of AI Cloud. The technology that transforms everyone to enjoy the benefits of expensive resources is the most meaningful technology for us, and will be remembered as the first step toward providing and democratizing AI."

 
▲ Director Jia Li gave a video lecture on AI innovation and Cloud.

 
▲ She said that there is a need to focus on key elements including computing, algorithms, data, talents and expertise.

 
▲ A variety of API are prepared for clients.

In the last step, Lee Seung-bae, CTO of Ticket Monster (TMON), took time to introduce cases of Korea partners.

According to the introduction, TMON now uses OCR technology among the Google Vision APIs to find words that should not be used in very small fonts in tens of thousands of product description images having thousands or tens of thousands of pixels in size. Moreover, the possibility of providing convenience services using speech APIs or natural language APIs is also under consideration.

Lee Seung-bae said, "Each of the techniques that uses machine learning is not good enough in terms of accuracy, but it could be a great choice for quick results."

 
▲ Lee Seung-bae, CTO of Ticket Monster, took time to introduce cases of Korea partners.


Copyright ⓒ Acrofan All Right Reserved

Contact : guide@acrofan.com / RSS
Copyright(c) ACROFAN All Right Reserved