Shulou(Shulou.com)11/24 Report--

Recently, Manfu Technology, a service provider of AI infrastructure and data intelligent platform, announced that it has completed tens of millions of yuan B round of financing in September 2023, and the current round of investors is ANPU Capital. The funds will be mainly used for AI infrastructure construction, closed-loop replacement of large model tagging platform and data tagging market expansion.

Manfu Technologies, which was commercialized in 2019, is a data-driven AI infrastructure platform company dedicated to gaining insights and value from data and building artificial intelligence applications in a more streamlined way to make AI lightweight and inclusive.

Its main products and services include: data intelligent platform for data life cycle management, AI data center, AutoLabeling platform, AutoML platform and basic data services (data tagging, data collection, data cleaning), etc.

With one-stop data solutions from strategy to technology landing, we have reached deep cooperation with hundreds of enterprises, including autopilot data tagging, AI data lifecycle management and so on. Users include mainframe factories, car-building new forces, front-line technology companies, mainstream algorithm companies and the world's top Tier1 manufacturers, etc., and the revenue is expected to achieve more than three times sustained growth in 2023.

Data definition model

After years of development, AI industry has gradually entered the intersection of technology and commerce.

The algorithm model has changed from the modeling improvement of increment to the iteration and optimization of performance to meet the more stringent requirements of model quality in business application scenarios.

Structured data has become the basis of AI algorithm model development and iteration. The strong "understanding" of AI is inseparable from the continuous input of structured data and the more refined use of data.

The AI industry is integrating around data-centric, and whoever owns the data has the right to define the model.

In the subdivision of the scene, the upsurge of self-driving city NOA is on the rise. Under the comprehensive innovation of the technical paradigm, the autopilot awareness algorithm has been upgraded to the BEV+Transformer architecture, and the end-to-end algorithm solution has become the mainstream, which promotes the autopilot awareness algorithm from lightweight CNN two-dimensional perception to the dimensionality enhancement based on Transformer four-dimensional perception, which also leads to an exponential increase in data demand.

Autopilot has reached a tipping point with the help of the AI model. But the precondition of quantitative change to qualitative change is the support of large-scale data-the qualitative change of Transformer model requires hundreds of millions of kilometers of labeled data, and covers the emerging Corner Case, which puts forward higher requirements for the scale of data production and the level of automation.

The ultimate goal of autopilot is to replace the driver, but before that, data markers need to be replaced by neural networks first.

AI-driven data Intelligent platform

As the BEV+Transformer technology route has become the core architecture of the new generation of autopilot perception capability, data closed-loop capability has replaced the algorithm paradigm and become the key to determine the outcome of commercial mass production from 1 to N.

Every step of the data closed loop is a game between cost and efficiency, and the data production capacity of low-cost AI becomes the key to boost the data flywheel.

As a leading AI infrastructure and data intelligent platform service provider, Manfu Technology takes product technology as the core driving force, and takes the lead in the industry to achieve low-cost, unlimited and large-scale production of AI data by building an AI+RPA-driven data platform to precipitate data Know-How capabilities.

The core product system of Manfu technology data platform is composed of MindFlow SEED data service platform and MindFlow AutoLabeling automatic marking platform. After many generations of version changes, it has established a technical barrier of 6-12 months in the field of 3D and 4D point cloud data processing.

In the specific application scene, the platform supports 2D, 3D, 4D full-category annotation in autopilot and other scenes, such as 2max 3D fusion, point cloud segmentation, point cloud sequence stacking, BEV annotation and so on.

In order to solve the problem of large-scale point cloud adaptation rendering in 4D point cloud tagging scene, Manfu Technology developed map slicing and LOD large-scale point cloud rendering technology, 4D point cloud lane and 4D point cloud segmentation scene can achieve the smooth operation of low configuration hundreds of millions of point clouds per frame.

4D point cloud segmentation and dimensioning scene

As an automated AI data platform, the building of RPA and AI capabilities is the core of Manfu Technology's construction of technical barriers.

The capability of RPA is mainly reflected in many aspects, such as process automation, scheduling and distribution automation and so on. On the other hand, the ability of AI has gone deep into all aspects of data flow, which is embodied in:

1) cover data preprocessing, algorithm inference to result refinement complete algorithm link, commercial static road adaptive segmentation, dynamic obstacle AI preprocessing, AI interactive tagging and other dozens of AI algorithm labeling models

2) using Backbone + multi-Head algorithm architecture to quickly adapt to different scenarios and greatly reduce the cost of research and development of multi-tasking model.

3) build AI tagging model based on AutoML and its own dataset, and complete algorithm iteration by self-driving.

4) using transfer learning, knowledge distillation and other methods, based on small batch data + bottom general large model fast output algorithm model.

The strong coupling of RPA and AI capabilities gives Manfu Technology the ability to provide more standardized data solutions with lower labor costs and marginal costs. The overall human efficiency is increased by an average of 80%, the data production cost is reduced by an average of 50%, and AI data is produced at a low cost, no upper limit and on a large scale.

Automatic tagging system of AI based on large Model

As a new base to lead the new round of leaping development of artificial intelligence, the large model is going deep into reality, empowering a thousand industries.

The capacity advantage brought by massive parameters gives the large model stronger performance and generalization ability, which provides a new technical solution paradigm for the traditional labor-dependent links such as data preprocessing and data labeling.

At present, Manfu Technology has completed the research and development of a large visual model marked with autopilot data. By introducing driving data to build RLHF, and building a large model based on deep learning and computer vision, we can achieve efficient data processing and automatic labeling in complex driving scenes.

The main technical features of the large model of Manfu Technology data labeling are as follows:

1) based on weak supervised and semi-supervised learning, efficient detection, segmentation and recognition of scene objects are realized through a small amount of manually labeled data and a large number of unlabeled data.

2) based on BEV multi-view fusion and 3D reconstruction, 3D information of scene objects is automatically generated with the help of multi-camera, lidar and other source data.

3) the transfer learning method is used to uniformly represent and learn the data in different scenarios and modes, so as to improve the generalization ability and adaptability of the model.

4) through feedback with manual labeling process, the performance of iterative model is continuously optimized by using active learning and interactive learning.

With the support of the above large model technology, the average efficiency of typical autopilot data labeling scenarios can be improved by more than 4-5 times, leading Manfu Technology to take the lead in stepping into the era of automatic data labeling.

Data-driven AI infrastructure

In the whole life cycle of algorithm update iteration, from design, training, evaluation to simulation, massive data input is needed as a support, in which data label is the foundation and starting point of the whole process.

If the Internet era is the handling of information traffic, then artificial intelligence, especially the large model era, is the handling and fine use of massive data.

On the basis of information flow, many far-reaching business models have been born. In the era of AI, anyone can use data as a "shovel" to explore the commercial "gold mine". Whether or not to have a higher quality and more quantity of "shovels" is the key to determine whether the "gold prospectors" can really take the initiative and find gold.

In the AI gold rush era, Manfu Technology hopes to play the role of serving the "gold prospectors", extending to the upstream and downstream of AI driven by data, and building a general AI infrastructure. Users can build AI applications and manage the whole lifecycle in a more streamlined way. At the same time, they can also adjust each component in the model building process more flexibly to get more consistent requirements and analysis results.

Under the guidance of the above vision, Manfu Technology has built an AI data closed-loop platform in self-driving scenarios, covering DaaS data tagging platform, data management platform, AutoLabeling platform and AutoML platform, providing end-to-end solutions from data preparation to model application, and extending to other AI application scenarios.

Among them, the data management platform integrates data storage, processing, import and export, and connects external platforms such as data acquisition platform, data tagging platform, model training platform and production and operation system through SDK, and enhances the functional experience of smart tags, analytical reports, scene mining and natural language search with the help of AI and big data technology, so as to improve the efficiency of data use and management.

The AutoML platform is an automatic training platform for general visual scenes such as autopilot, which provides automatic training and fast iteration of algorithm models. It can automatically optimize the model structure, parameters and super parameters, improve the performance and generalization ability, and realize zero code one-button training, unattended.

Manfu Technology AI Infrastructure

Manfu Technology AI infrastructure solutions cover the data layer to the algorithm layer, providing both DaaS and MaaS services. Regardless of user size and AI R & D capabilities, as long as there is AI demand, you can use the infrastructure provided by Manfu Technology to easily create proprietary AI products to achieve the transition from data to business value.

AI For Everyone

The independent data closed-loop platform, AI data production capacity and keen insight into customers' business needs enable Manfu Technology to achieve higher-than-expected business growth in the ever-changing market, and the data-driven AI infrastructure validates application value and commercial potential in industries such as self-driving.

In the next stage, Manfu Technology will continue to dig deep into the data industry and constantly improve AI infrastructure construction. Just as AWS is to cloud computing and Snowflake is to data analysis, Manfu Technology hopes to build a common infrastructure based on data in the AI industry to help users train and deploy artificial intelligence applications in a more streamlined way. Whether they are startups, mature companies or individuals, they can enjoy the convenience of AI with a simple click or a few lines of code to achieve the true democratization and inclusiveness of AI.

In the Internet era, Google controls the entrance of Internet traffic with its search engine, and Microsoft controls the upper reaches of the PC ecological chain by virtue of its operating system. There are no trillions of tech upstarts challenging Google, Microsoft and so on, but AI is making everything possible.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.