One sentence generates a digital human image, Kunlun Core 2 goes into mass production... Baidu Brain upgrades to 7.0, Wang Haifeng: The technology is stronger but the threshold is lower

Mengchen Yuyang sent this from Aofei Temple
Quantum Bit Report | Public Account QbitAI

Baidu and CCTV have teamed up again to showcase black technology during live broadcasts.

Baidu CTO Wang Haifeng only said one sentence and gave the host a digital "twin brother".

What’s even more amazing is that as soon as this digital man was born, he started speaking with a tone that sounded like “Peking University is okay” (dog head).

This is not the end. At the scene, Zhurong, the digital human Mars rover jointly developed by China's Mars Exploration Project and Baidu, also met with people across the country.

As soon as he came on stage, he chatted and laughed with the host in a calm and relaxed manner, and he came up with all kinds of jokes such as "admission by recommendation" and "planting potatoes on Mars" with ease.

When talking about something exciting, he even improvised a poem, which made the host exclaim "Look at how good you are".

This is the live broadcast of the Baidu World Conference, another technological feast contributed by Baidu on stage.

Speaking of the highlights of this live broadcast, it’s not just the digital human, but also a host of new products and highlights such as car robots, Kunlun Core 2, Xiaodu smart giant-screen TV, etc.

Of course, behind all the excitement, there has always been a technical foundation to support it. Moreover, as a company that has been deeply engaged in AI technology for more than ten years, Baidu's in-depth layout and latest technical insights displayed through the Baidu World Conference are often more worthy of attention.

Baidu Brain upgrade 7.0: integrating innovation and lowering barriers

For example, the digital person who chatted with the host this time was actually supported by the new capabilities of Baidu Brain .

Wang Haifeng introduced that this Baidu AI core technology engine has now been officially upgraded to 7.0.

The so-called "upgrade" not only adds functions and enhances performance on the surface, but also targets the two major trends in the development of the AI ​​industry that Wang Haifeng has previously identified - integrated innovation and lowering the threshold .

Why is fusion innovation the trend of AI development?

Today, as artificial intelligence enters the stage of industrialized mass production, a single breakthrough in a specific technology is certainly a good thing, as it can lead to the publication of papers, rankings, and awards. However, if we want to truly enter industrial practice and implement it, it is no longer feasible without the integration of multiple technologies.

Wang Haifeng pointed out four specific integration directions for this phenomenon:

• Fusion of knowledge and deep learning

• Cross-modal multi-technology integration

• Integration of technology and scenarios

• Software and hardware integration

The reasons specifically correspond to the four current situations of AI development today, from technology to application.

The first current situation is that AI has indeed achieved miracles in recent years, but this great effort comes at a high price.

The GPT-3 with hundreds of billions of parameters has astonished the world once in terms of language ability, and the cost of training it once again shocked the world by consuming about 190,000 kWh of electricity. How many companies can afford to use AI applications developed based on such high-cost technology?

Therefore, after GPT-3, the development of language models split into two routes.

One is to continue to increase the scale, represented by Google's Switch Transformer, which has pushed the parameter scale to the trillion level in one fell swoop .

The other is that Baidu Brain 7.0 integrates knowledge and deep learning , trying to let AI master more common sense beyond large-scale computing.

The knowledge-enhanced large model released by Baidu Brain 7.0 , by introducing a large-scale knowledge graph, only uses tens of billions of parameters to top the global list of the authoritative language model evaluation SuperGlue, surpassing the human level by 0.8 percentage points.

Baidu is able to rely on knowledge instead of brute force to create miracles, and it also has a unique advantage: it understands Chinese better and has the world's largest heterogeneous knowledge graph.

In addition to topping the international rankings, this knowledge-enhanced model also refreshed 54 Chinese NLP task benchmarks in one fell swoop.

After learning knowledge such as "Sa Beining is the host" and "No water has been found on Mars", Zhurong digital talents were able to answer questions accurately and have fluent conversations at the conference.

Baidu Brain 7.0, which has such capabilities, has lower costs, stronger capabilities, and is closer to the market, and can really start talking about large-scale implementation.

The second reality is that for AI to be implemented, it cannot just provide some fragmented technical support behind the scenes, but must start to interact more directly with humans.

If you want to interact with people, understanding and generating text is only one aspect. To see, listen, and speak to people, you need to integrate text, voice, and image multimodal technologies.

The "one sentence to generate a digital human image" demonstrated in the live broadcast is a wonderful demonstration of the cross-modal multi-technology integration of Baidu Brain 7.0 .

First, you need to listen to the instructions and convert them into text that is easy to calculate. Speech recognition technology is used here.

Understanding the meaning of instructions and determining what kind of digital human to generate involves natural language understanding.

Then, we execute the instructions, extract the corresponding visual features and generate the corresponding model, relying on a series of computer vision technologies.

With cross-modal fusion like this, digital humans can finally appear on CCTV in a complete form, giving the audience a unique technical science experience.

Of course, showing the charm and value of technology to the public is only a small part of what AI needs to do. For AI, it is more important to implement it in all walks of life and do practical things.

The third current situation is that if AI is to be implemented, it must truly demonstrate its effectiveness and generate benefits in real business scenarios. Otherwise, who will pay for AI?

One of the characteristics of AI at this stage is that it is only good at solving specific tasks. Taking large-scale pre-training models as an example, in the pre-training-fine-tuning paradigm, the most important thing is to find the right pre-training task.

Like Baidu's previous ERNIE2.0 language model, it relied on specifying tasks at three levels: words, sentence structure, and semantics, and combined with innovative continuous multi-task learning. It surpassed the then-powerful BERT on the authoritative test GLUE, laying the foundation for the subsequent knowledge-enhanced large model.

Although what is discussed here is a purely technical task, if we change our thinking and regard running tests and rushing to the list as the business scenario that AI companies have to face, then if AI is to be implemented in more industries, the answer to where to find tasks is obvious:

Go to the scene and integrate technology with real business scenarios .

For example, at the Baidu World Conference this time, the conversation between the host and the guests was recorded by the Ruliu Intelligent Conference Minutes System. Relying on AI technology, in addition to voice recognition recording, it can also automatically extract important content to form a summary.

The high-quality, low-latency simultaneous interpretation effect demonstrated by the on-site simultaneous interpretation system is the result of the integration of Baidu Brain 7.0 machine translation technology and simultaneous interpretation scenarios.

The last current situation is that the actual implementation of AI in business scenarios is not only a matter of technical development, but also requires consideration of deployment.

With the algorithm, the next step is to think about how to use computing power more efficiently. This requires the integration of software and hardware . Apple, for example, has taken the path of deeply integrating algorithms and chips to make mobile phones play the role of computers.

In terms of hardware, Baidu this time brought out its self-developed AI chip Kunlun 2nd generation, which uses the world's leading 7nm process and Baidu's self-developed 2nd generation XPU architecture, and has been mass-produced.

In addition to having a performance that is 2-3 times higher than the previous generation, the Kunlun Core 2nd generation can be applied to a variety of scenarios such as cloud, terminal, and edge, thus covering many fields such as Internet core algorithms, smart cities, and smart industries. It also has the potential to be used in cutting-edge research such as high-performance computer clusters, biological computing, intelligent transportation, and unmanned driving.

Like this, Baidu Brain 7.0 unleashes more possibilities through the integration of software and hardware .

As Baidu Brain continues to upgrade in the above four integration directions, the AI ​​technologies involved are becoming more and more complex.

However, the scope of AI services is getting wider and wider, expanding from technology companies and Internet companies familiar with AI to traditional industries, public utilities and personal life.

Lowering the application threshold has become another major trend in the AI ​​industry

Wang Haifeng said at the Baidu World Conference:

Baidu's AI technology is getting stronger and stronger, but the threshold for application has been lowered.

How is it done?

This is thanks to the core foundation of Baidu Brain - the PaddlePaddle deep learning platform.

Baidu PaddlePaddle is open to the whole society as an open source platform, and this whole society is not just talk.

For professional developers, Baidu PaddlePaddle provides an industrial-grade open source model library, large-scale distributed training technology, and a high-performance inference engine that can be deployed on multiple terminals and platforms.

For developers in their growth stage, Baidu PaddlePaddle also has the AI ​​Studio learning and training community, which provides an online programming environment, free GPU computing power, open source data and algorithms, as well as AI learning paths and various competitions to guide developers in their learning and growth.

For entry-level developers, Baidu PaddlePaddle also has the EasyDL zero-threshold AI development platform, which allows developers to develop their own AI applications even if they don’t know code.

So far, more than 3.6 million developers from all walks of life have developed 400,000 AI models through the PaddlePaddle platform , serving a total of 130,000 enterprises and institutions , covering industries, quality inspection, agriculture, medical care, urban management, transportation, finance, sports and other fields.

In addition, Baidu PaddlePaddle has also cooperated with many universities to organize deep learning teacher training for teachers, hosted a number of events such as the China University Computer Competition, and provided internship programs and employment guidance for university students.

At this Baidu World Conference, Wang Haifeng also announced the establishment of Baidu Pinecone Academy, which will gradually implement its goal of cultivating 5 million AI talents in the next five years by providing basic courses, technical competitions, industrial training, and scientific research funds.

On the other hand, cultivating talents and improving their capabilities can also lower the application threshold for them when they enter the industry in the future.

In the end, it is not a bad idea to regard lowering the threshold as another form of integrated innovation in ecosystem construction.

Technological innovation + industrial development = new driving force for economic development

Having said so much, it is not difficult to summarize Baidu’s latest technical thinking behind the upgrade of Baidu Brain.

The core is that to further stimulate the value of AI technology, it is necessary to organically integrate technological innovation and industrial development.

Nowadays, data resources are already a recognized key production factor. Therefore, based on the technical foundation of Baidu Brain, Baidu Smart Cloud proposed the strategy of "digital transformation + intelligent upgrading" to help the industry become intelligent in one step, which is actually in line with the latest trend of the development of the digital economy.

Simply put, in traditional industries such as finance, medical care, manufacturing, energy, and even entertainment and sports, it is necessary not only to build automated systems to improve production efficiency, but also to make full use of one's own digital assets and use intelligence to stimulate new growth points.

Baidu has also provided actual examples of how to do it.

Behind the Chinese diving team, which just won 7 golds and 5 silvers at the Tokyo Olympics, there is an invisible "AI coach".

At the Baidu World Conference, Baidu founder, chairman and CEO Robin Li and Chinese Diving Association Chairman Zhou Jihong had a live chat. Chairman Zhou shared how the "3D+AI" diving training system provided by Baidu Smart Cloud has changed daily training:

The problem of diving data collection and analysis has been solved. With this system, feedback can be seen in 3 seconds. Sometimes when I am out for a meeting or a business trip, I can also see the training situation of the team, and the feedback is very timely. The system can cut different angles and dig out the details of different movements, which are clear at a glance.

This is also verified in areas closer to daily life.

For example, in Quanzhou Water Group, Baidu Smart Cloud and Aridi jointly built the Quanzhou Water Brain based on Baidu Brain, which achieved automated control and operation and maintenance of the production process, greatly reducing the pressure of personnel management.

At State Grid Xinjiang Power, based on the "AI middle platform" jointly built by Baidu Smart Cloud and State Grid Power, power station inspection robots can replace humans to capture every subtle error in the inspection process, which not only alleviates the problems of shortage of inspection manpower and harsh working environment in remote areas, but also provides strong guarantees for the west-to-east power transmission.

In the famous tourist city of Lijiang, the "City Brain" built by Baidu Smart Cloud is committed to making residents and tourists live in harmony: detecting the cleanliness of the environment and ensuring the safety of tourists through an emergency plan system...

There is also the "smart factory" created by Baidu Smart Cloud and Suzhou Industrial Park. In the Kaibo workshop in Changshu, Suzhou, AI can help quality inspection workers "listen to sounds" and "recognize images" to improve manufacturing efficiency and quality. It can also process manufacturing experience such as operation manuals and maintenance history records in a knowledgeable way, promoting China's manufacturing industry from "manufacturing" to "smart manufacturing".

In fact, this is just a typical example. Cases like this can be said to be happening every day all over the country...

Ultimately, this is the deeper meaning and thinking behind "integrated innovation" and "lowering the threshold": such a development path is actually about how to better combine technology itself with industrial development on the basis of technological innovation to form new impetus for economic development.

It is extremely exciting to see what new achievements this wave will drive China's digital economy to create.

What do you think?