Alibaba’s thorny road to artificial intelligence: What does it feel like to sit on a mountain of gold and eat steamed buns?

Going online does not mean the end of the challenge. For example, there are still unfathomable entrances. Pan Pan clearly remembers that the first time the image search was launched was in a four-level menu of Taobao Mobile. "Fourth level" means that you first find "Discover" on the homepage, then click on "Featured Services", click "More", and then...

As Alibaba's most core business battlefield, Taobao Mobile's attitude is obvious - "willing to give technology opportunities, but also requires that risks be controllable."

Opportunities need to be fought for, and more importantly, their value needs to be "proven" by oneself.

From the initial thousands of daily active users, to over a million, over ten million, and finally to over 20 million, the image search application has leapt all the way to the top of the shopping guide category on Taobao's homepage. But at the same time, there are constant doubts from the outside world. "I remember very clearly that every year everyone asks, can the data still grow? ... Are you still doing it? What are you doing?" Pan Pan said.

Not being understood seems to be the fate of pioneers. Qi Yuan recalled the early period of exploration, "The biggest training for me at that time was that it is normal not to be understood."

This may be called a kind of optimism, but there is no doubt that belief in technology is an indispensable solid force when facing difficulties and misunderstandings.

After an anxious and difficult "struggle" for resources, Qi Yuan's distributed machine learning platform was finally launched. In order to launch the platform as soon as possible, the team gave up their year-end vacation and stayed to work hard during the Spring Festival. The backbone members held intensive meetings and brainstorming. "At that time, I felt the team's fighting power firsthand. Everyone really believed that as long as it was launched, the effect would be good," Qi Yuan recalled.

During the Double 11 Shopping Festival, the platform realized the large-scale application of personalized recommendations on Taobao and Tmall for the first time. At the Alibaba Group Algorithm Awards that year, Qi Yuan led a team of 80 people to win 6 out of 16 awards. Now, as the head of the financial intelligence direction of the DAMO Academy, Qi Yuan leads the team to build intelligent analysis and decision-making technologies for financial and economic scenarios.

On the other hand, after upgrading search categories, optimizing data, and iterating algorithms, the “Image Search” project completed three entry transitions and finally entered the homepage during the Double Eleven shopping festival in 2015. What makes Pan Pan very proud is that the data of Image Search has always relied on natural growth, and almost no commercial promotion resources of Taobao Mobile have been used.

The essence of portal upgrade is a process of continuous "self-proving", which is completed by technology and the promoters behind the technology through repeated challenges, and then solidified into an iterative tradition of Alibaba's technology products.

Different from the public reasoning and simulation verification in the ivory tower and in front of the test bench, "self-proven" in the business field requires real business indicators and actual results, just like the slogan that can be seen everywhere in Alibaba's office area - "NO DATA, NO BB".

Behind the rapid growth is the major system overhaul every two years. "Alibaba is actually very particular about innovation, and we generally don't rehash old ideas," said Pan Pan, who divided the development of the graph search system into three periods. "Each upgrade is not only an improvement in the algorithm, but also an improvement in the overall thinking."
"Data, system, and algorithm are interdependent. Different perspectives on data cognition and processing methods have given rise to matching algorithms and engineering systems, so upgrades are at the level of the entire system."

In the early stage of the project, the amount of data was small and manual annotation was required, so we studied the system and algorithm of the small model to match it; then the training data was liberated, and the team tried to analyze the logical relationship between data and ranking through three types of user behavior data (query data, click data, and non-click data), and generated a triple Deep Ranking framework, and the corresponding training framework and system were upgraded and iterated; last year, Graph Search began to connect to the ultra-large-scale parallel processing platform, releasing the training capacity of billions of data.

Unlike other image search applications on the market, such as Baidu Image Recognition, WeChat Scan, and JD Photo Purchase, Alibaba emphasizes "universal" capabilities. For example, it can not only support all physical product searches on Taobao Mobile, but also include the recognition and classification of non-products such as QR codes, plants, and garbage. These functions are all integrated into one image search window, without the need for a second jump, and the average daily active users have reached more than 20 million.

On the growth path of researchers, experiencing a complete technical engineering implementation is of great significance. It not only hones practical capabilities, but also provides a window for them to gain an in-depth understanding of what the business wants and needs.

"This experience has helped us understand what kind of technology we should create and promote, which technologies are more likely to become products, and how to make a product effectively support the business," Jin Rong said.

It can be said that the practical accumulation from R&D to business is the cornerstone of Alibaba's creation of an AI technology implementation system.

Today, this set of ideas is implemented in Alibaba's technical thinking, big and small. For example, in the speech synthesis technology KAN-TTS (Knowledge-Aware Neural TTS) launched by the speech technology team this year, the team considered the model deployment environment in different environments in advance, and carried out framework design and efficiency optimization. It improved more than 20 key algorithms based on various needs, and finally achieved the fastest deployment in the cloud or terminal, even in the case of limited CPU and storage.

Whether a technology can take into account the various practical problems encountered in the project deployment phase at the beginning of its design is an important sign of the maturity of the technology implementation system.

3. The China Research Institute has no Benchmark
"It is the technological lifeline of Alibaba's international business. Without these capabilities, it would be difficult for Alibaba to be called an international company," said Si Luo with great confidence.

He is the head of the Language Technology Laboratory of Alibaba DAMO Academy, leading a team of more than 100 people to develop underlying technologies such as natural language understanding, machine translation, and cognitive intelligence. These technologies are hailed as the crown jewels of artificial intelligence.

Si Luo, currently the head of the Language Technology Laboratory at DAMO Academy

Si Luo focuses on research in the fields of machine learning and NLP. After becoming a tenured professor in the Department of Computer Science at Purdue University in 2012, he established his position in the academic circle. Si Luo joined iDST in 2014, along with Jin Rong and Qi Yuan, and was an early member of Alibaba's artificial intelligence technology system.

Compared with vision and voice technologies that are closer to front-end users, language is more bottom-level and functions in the form of atomic capabilities, playing the role of empowerment and support. Because of this feature, it is often indispensable for large Internet technology companies, but the technical teams are extremely scattered.

In that case, why not concentrate our efforts to open it up?

Therefore, language technology happens to be the best testing ground for technology platformization.

October 2016 was an important time for Si Luo and the NLP team. Before that, they were busy taking on one "project" after another, and had participated in projects such as "Juhuasuan", "AIios" and "Taobao Toutiao".

After that, Si Luo was given the task of establishing the "big middle platform" of NLP. In other words, the NLP of each business line of Alibaba would no longer operate independently.

In order to allow other business lines to access the platform, Si Luo adopted the "brand effect" approach.

Yes, it is also necessary to build a brand within the company, especially in the process of building a scalable platform.

"First, we need to provide very personal services, then let key top users use it, gradually let them realize the benefits of NLP platforms and solutions, and then spread the value of our technology through seed users," Si Luo shared the "routine" for creating brand effect.

It is precisely because of the emergence of technology platform that this team of more than 100 people is able to support more than 600 business parties of Alibaba, with the daily call volume reaching 2 trillion times.

This is a leap from quantitative change to qualitative change in the influence of NLP technology.

Following the same idea, another technical branch led by Si Luo - machine translation technology has also achieved large-scale business support capabilities, providing machine translation services in more than 20 languages ​​and 48 language directions to buyers and sellers on Alibaba's global e-commerce platform, covering the vast majority of countries in Europe, Asia, America and the Middle East.

After the upgrade of Alibaba's image search application to a technology platform, it has become the world's largest image search application system, from the earliest support for the Taobao mobile platform to the current landing in the six mainstream apps of the Taobao system. "This is determined by the market share of the Taobao system business," said Pan Pan.
From a single functional application to the support of ten, one hundred, and hundreds of business parties, it is used by hundreds of millions of users every day. The strategy of technology platformization has been effectively verified and recognized. Especially for technical personnel, through the process of platformization, the influence of single-point technology has been continuously generalized, and the brand effect of technology has been continuously amplified.