2023 The four major AI directions that need to be focused on

In the past 10 years, artificial intelligence has moved from the laboratory to all walks of life, becoming a common technology application in many fields. According to IDC data, the global AI market size reached 432.8 billion US dollars in 2022, an increase of nearly 20%. The PRECEDENCE Research predicts that by 2030, the artificial intelligence market will exceed $ 15.7 trillion. This explosive growth is based on the reduction of data acquisition and storage costs and semiconductor innovation. Heterogeneous component integration simplification

How to perform more artificial intelligence treatment on marginal nodes is one of the important directions for the efforts of AI in recent years. Fortunately, many valuable innovations have been produced in this direction. This year’s direction is still the key direction we need to pay attention to, but the focus will turn to integration and landing.

For marginal devices, the artificial intelligence system needs to perform various tasks. This requires not only different types of computing power but also different types of memory, connection, and sensor input. How to integrate these heterogeneous components efficiently, energy saving and stability is a huge challenge. This work involves other engineering fields, such as mechanical design, optical design, electrical design, and digital and simulation semiconductor design.

The integration of data centers has brought a series of different but very diverse challenges. In order to provide the computing power performance required for in-depth calculations, multiple cores need to be combined to one chip. These components are mainly composed of dense digital logic, such as accelerating the basic computing power required for large neural networks.

Artificial intelligence design tool
With the continuous increase of chips and process complexity, the adoption of artificial intelligence design tools has grown in index level. In just one year, the number of commercial chips designed by artificial intelligence has increased at least one magnitude. With the accelerated development of artificial intelligence design technology, training data sets have become more comprehensive, and new tools will bring more advantages to designing teams.

With the maturity of this technology in the next year, the design capabilities driven by artificial intelligence will achieve productivity breakthroughs in the chip design field and help create more complex designs to meet the power, performance, and area needs. This year, there will definitely be applications based on AI-strengthening learning to solve various design challenges to enter the market. Even as the technology matures in the future, artificial intelligence design may become the mainstream of the market.

Generate AI
The development of new artificial intelligence applications is the most challenging and most time-consuming job to model, and then train and optimize it to perform specific tasks. This has given more research on the so-called basic model (also known as the big model).

The basic model is an artificial intelligence model. It only needs to be designed once and then uses a very large data set to train to achieve various goals. Once the training is completed, this model can adapt to many different applications, thereby reducing the time for designing new models for each application. The huge scale of basic models allows users to achieve new features that could not be realized before.

The basic model also promoted the very popular AI technology last year -generating AI. AI focuses on creating new content. The essence of its essence is the basic model, which can train very large data, including text, images, voice, and even 3D signals from sensors. According to the input, you can train the same basic models to synthesize new content, such as painting, creating music, and even creating chat robots. ChatGPT, a fire at the end of last year, is a typical application of basic models and generating AI.

Generating AI will make the creation of new content extremely easy and have extremely high application value. With the maturity of technology, coupled with the ChatGPT detonation market, I believe that this year’s generation of AI will grow.

Causal AI
Finally, I will continue to pay attention to cause and effect AI. One of the dilemmas of deep learning is the irreplaceability of the model. One way to solve this problem is to introduce causality. Although cause and effect AI is still early, it has been widely adopted in specific fields such as medical and health, drug research and development, financial services, manufacturing and supply chain organizations. By combining knowledge diagrams with causal maps and simulation, thereby transcending relying on historical data -based machine learning. Causal predictions can improve the ability to interpret artificial intelligence by making causality transparent.

http://oaicon.com/index.php/2023/02/25/2023-the-four-major-ai-directions-that-need-to-be-focused-on/

On the other hand, an important bottleneck that hinders the development of artificial intelligence is the lack of high-quality label data. Although we have seen progress today, we will not be fundamentally resolved in 2023. There are currently two directions to solve this problem. One is self-supervised lead, and the other is cause-and-effect AI. Self-supervision, and learning use self-supervision algorithms to conduct pre-training models, and then fine-tune the model based on specific tasks; causality AI does not require big data samples. The best and most effective example in this regard is NLP (natural language processing), where mask language modeling (Masked Language Modeling, which predicts the hidden words in the sentence) and causal language modeling (Causal Language Modeling, let the model predict sentences The next word) technology completely changed the traditional algorithm.