In his latest research, Tian Feng, a senior researcher at SenseTime, has successfully achieved in-depth insight into the changes in the earth's resources by using the remote sensing model "Earth Boundary". The application of this innovative model provides strong technical support for resource monitoring, environmental protection and other fields, and helps to understand the distribution and change of the earth's resources more accurately.
The 2023 EmbeddedaiSummit was successfully held in Shanghai on December 16, 2023. The conference was organized by ACMSIGBEDCHINA and supported by the Department of Computing of Harbin Institute of Technology, Tsinghua University, Shanghai Jiao Tong University, Shanghai Artificial Intelligence Industry Association, and Shanghai Zhangjiang Open Innovation Center, Institute of Artificial Intelligence, Harbin Institute of Technology. Embedded intelligence technology is an implementation of artificial intelligence applied to embedded systems, which can be applied to various smart devices such as smart cameras, smart speakers, and smart sensors. The EmbedDedaiSummit is a platform that aims to promote academic research and engineering technology exchange in the field of embedded intelligence, providing a place for relevant scholars and engineers to share new achievements, new technologies and new products. At this conference, Tian Feng, President of SenseTime Intelligent Industry Research Institute, delivered a special speech entitled "Large Model: The Road to the Future", which in-depth the development trend and future prospects of artificial intelligence large models.
Picture: Tian Feng, President of SenseTime Intelligent Industry Research Institute, gave a speech.
The "cost demand curve" of new technologies: new technologies continue to reduce costs, and market demand expands exponentially.
Only by continuing to reduce the cost can we achieve the explosion of social-level demand, just as the reduction of electricity has brought prosperity to the electric age, and the reduction of communication has brought prosperity to the information age, then with the continuous reduction of the cost of data, computing power and models, we will usher in the AI era, and the environment in which everyone can start an AIGC business will be launched, and artificial intelligence products that everyone can afford and use well.
As Xu Li, chairman and CEO of SenseTime, once said: "Whether AI can be used to name the era at this point in time depends on whether it can reduce the cost of production factors in our era on a large scale, so that AI can enter thousands of households." ”
Super Moore's Law: The demand for computing power of the model far exceeds Moore's Law.
As AI models evolve, the complexity and scale of the models continue to increase. Large, complex neural networks demand more computing resources and computing power, which goes beyond Moore's Law's traditional expectation of increased computing performance, and the "new Moore's Law" emerges.
The computing power required for early AI models doubled every 21 3 months;After the advent of deep learning in 2010, the model's demand for AI computing power doubled in 5 to 7 monthsIn the era of large models, the AI computing power required for large models will be shortened to double every 1 or 2 months by 2023, far exceeding the growth rate of Moore's Law.
As a result, there is a phenomenon of "computing power hunger", which is manifested in the profit margin of GPU chip manufacturers exceeding 1000, while AIGC application innovators are facing the dilemma of losing money.
The expensive first-generation AI chips will be replaced by more advanced and lower-priced new products, otherwise it will inhibit the wave of AIGC innovation and entrepreneurship for a long time.
SenseTime AIDC (large device): Let the domestic GPU better adapt to the domestic large model.
SenseTime Shanghai Lingang AIDC is currently one of the largest artificial intelligence intelligent computing centers in China, providing a large-scale application environment for domestic smart chips and forming more cost-effective AI computing services.
Through the "Artificial Intelligence Computing Industry Ecological Alliance" led by SenseTime, AIDC will accelerate the innovation of domestic smart chips and bring supporting support from scenario demand traction to the whole chain of software value ecology. It will accelerate the tackling of the next generation of general artificial intelligence technology and promote the innovation of domestic computing power chips for artificial intelligence.
As of the end of August 2023, the number of GPUs on the shelves has reached about 30,000, and the high-quality training data of 2 trillion tokens per month has been produced, and it is expected to exceed 10 trillion tokens by the end of 2023.
With abundant and leading computing power and data resources, SenseTime continuously optimizes and iterates its large model capabilities, and at the same time creates a leading large model landing and generative AI application ecosystem, and promotes more industrial intelligent upgrades by helping customers build industry large models and more field applications.
Will the high-quality texts available for training in human history be "exhausted"?
The demand for data for large models is skyrocketing, and the high-quality text that has historically been available for training will be "exhausted" by 2026.
We will run out of low-quality linguistic data by 2030-2050, high-quality linguistic data by 2026, and visual data by 2030-2060. This could slow down the progress of machine learning. ”
This means moving from computation being the main bottleneck for the growth of machine learning models to a pattern where data becomes a strict constraint.
If data lack is to become a bigger problem in the future, we need to be more progressive in the field of data.
For example, unlabeled data has never been a constraint in the past, so it is possible to produce more results in terms of the efficiency of unlabeled data;There are also more possibilities in terms of high-quality data, where high-quality data can be extracted from low-quality**.
Just like the SenseTime remote sensing model "SenseTime", it can analyze remote sensing data at any time and resolution in China, provide a complete set of structured vector data, and create a complete chain from remote sensing data acquisition, data analysis to application of results, providing a steady stream of high-quality data.
AI Remote Sensing, Insight into the Mystery of the Earth: SenseTime's "Earth Boundary" AI Remote Sensing Model.
Based on the general visual model, SenseTime's "Ground Boundary" AI remote sensing model uses 70 million annotated remote sensing samples, which is a large remote sensing vertical model. In order to balance the inference speed and interpretation effect of the interpretation platform, 3.5 billion parameters were used.
In the three mainstream tasks of remote sensing, there are the best industry performance. At the same time, in order to ensure the generalization ability of the large model of the geographical boundary, we have adapted the mainstream remote sensing satellite data sources, with a resolution from 0 5m to 10m, and also maintain the spatiotemporal coverage of typical global landforms in terms of seasons and landforms.
At the same time, we have introduced generative capabilities to ensure that the remote sensing interpretation results are in line with manual mapping planning.
Figure: SenseTime's "Ground Boundary" AI remote sensing model.
Technological breakthrough of the "Boundary" remote sensing model: senseearth3 0 released 51 semantic segmentation models in the SenseTime remote sensing model, which used to take 1 unit time to run 1 category, but now only 1 unit time can be used to complete the interpretation of multi-category data, which greatly saves users' time and costs.
In addition, the platform has also released 5 types of target monitoring, 4 types of change detection and 2 types of super-resolution algorithms. At the same time, it has also achieved a breakthrough in interpretation accuracy, in which the average accuracy of the feature segmentation capability on the million-level patch verification set exceeds 80, which can directly meet the application requirements of various business scenarios.
AI remote sensing empowers all industries: Based on the AI remote sensing model, SenseTime's remote sensing business has served more than 20,000 industry users, covering natural resources, agriculture, finance, environmental protection, photovoltaic and other industries.
Especially in the field of natural resources, SenseTime's AI general change detection has been widely used in natural resources law enforcement and supervision in more than 14 provinces and cities by virtue of its capabilities, helping users improve their work efficiency by 3 to 5 times.
Figure: AI remote sensing empowers all industries.
Typical case of AI remote sensing: a municipal big data center and a "boundary" remote sensing large model.
1) Rice identification.
A municipal big data center used the crop identification ability of the "land boundary" remote sensing model to detect the distribution of rice planting in the key monitoring areas of agricultural planting, and found the unreported rice planting areas, which provided objective and valuable reference data for understanding the real rice planting area and distribution in the region and supporting regional agricultural production management, which was highly praised by the leaders of relevant regulatory departments.
Figure: Typical case of AI remote sensing: rice identification.
2) Construction site inspection.
A city used the building change detection ability of the "land boundary" remote sensing model to detect the distribution of construction sites in the city, and detected nearly 20,000 construction sites in the city in only 6 2 hours, providing timely and objective data for the decision-making and guidance of city leaders.
Figure: A typical case of AI remote sensing: building change detection in a city.
3) "Waste-free city" solid waste testing project.
A city plans to carry out remote sensing detection of solid waste by using the ability of the "land boundary" remote sensing model, providing intelligence support for the regional reduction of solid waste generation and random stacking, and helping the region to comprehensively promote the construction of a "zero-waste city".
Figure: Typical case of AI remote sensing: "zero-waste city" solid waste detection project.