从Web2到Web3：我为何看好AI赛道

Author Jingwei China investors from the original Twitter I recently began to focus on spending time to look at the track in the global model big track, the open source model community data big track serves all kinds of middleware of the big model, such as the whole process service and part of it to become a big industry model. We welcome all kinds of entrepreneurs to communicate with Jingwei. We think it will be a long-term track. The first phase will share the data we have laid out recently. The data labeling industry in the big track is also my personal investment this year. The development of satisfactory targets can be divided into data preparation with data collection, cleaning, labeling and enhancement as the main body, and algorithm development with model construction, training, optimization and deployment as the main body. Among them, because of the diversified demand for data in the new era, data in the new era is also highly dependent on human labor, and it also needs to be further improved and people's smooth interaction to increase efficiency. Data labeling refers to the characteristics of data samples needed for model training Because the current development is still in the stage of supervised learning, the algorithm model represented by deep learning learns and verifies the connotation information of data and the logic between data in the training process. It is necessary to realize data labeling based on data feature identification, which is one of the core tasks of data preparation and even artificial intelligence project development. Similar to the other workflows of data preparation, data labeling is highly dependent on human labor, lengthy work cycle and huge labor cost. It has become one of the main factors restricting the development of the industry. The pain point on the supply side of data labeling service has given birth to the demand for automation tools in the market, which has promoted the development and large-scale application of intelligent data labeling technology. From data collection to available data sets, in the downstream intelligent driving field, it still needs a lot of manpower to label various scenes, such as cats, dogs, telephone poles and strollers. For example, they are important data providers who have established their own numbers in third world countries around the world. According to the annotation studio, the data annotation of characters and pictures is assisted. However, with the progress, the proportion of pre-annotation in the workflow is gradually increasing. In the early days, data annotation was mainly done manually to build and accumulate the data set of machine learning. Although the efficiency is relatively low and the cost is high, the data provided to the machine has great advantages as long as the annotation is in place. With the passage of time, the focus of manual annotation has gradually shifted from the United States to third world countries such as Venezuela and the Philippines to reduce the cost. The development of automatic data labeling can improve the accuracy of data labeling. For example, the model can be used to assist manual labeling, such as preprocessing the data and then sending it to manual labeling, or the labeling results provided by the automatic model can be manually reviewed and corrected. Compared with pure manual labeling, auxiliary labeling accelerates the speed of data labeling. At present, one of the largest data labeling companies in the world is trying to reduce the proportion of manual participation in the data labeling process, although pre-labeling has been achieved in the field of computer vision. Good results, but in the new era of language and large model, pre-labeling is still very immature and cannot completely replace manpower for the following reasons: low accuracy, especially when dealing with complex tasks and marginal cases, sample deviation and model illusion. Some vertical fields need a large number of data sets labeled by industry experts, and the scalability of pre-labeling is poor, especially for small languages or uncommon scenes, which are costly and of poor quality, and still need specific manual completion. To sum up, pre-labeling will not completely replace manual labeling in the short term. Note that the two will coexist. Although the proportion of manual labeling may decrease, auditors are still needed to review the data labeling in the labeling process. The data labeling process under pre-labeling is not a new industry. With the rise of intelligent driving, the data labeling industry began to emerge in. The following figure shows the market size of domestic predicted data labeling providers. It is worth mentioning that the data labeling market in the United States is roughly twice as large as that in China. The data labeling industry is a relatively scattered market, unlike a technology. The field with extremely high technical barriers is more like a field where technical manpower and organizational management barriers each account for one third. The core competitiveness of this field is mainly reflected in the following aspects: price, quality, professional knowledge and knowledge coverage, diversity, speed and price. Because everyone needs a lot of cheap data, the pressure on price drives a geographical arbitrage way, that is, in the developed United States, it may cost US dollars to complete a data annotation, while in the less developed China, it is obvious. Only US dollars is needed. In the Philippines, it may only cost US dollars. Therefore, one of the solutions in the market is to hand over the order to the first world countries and then recruit people in the third world countries to solve the problem through direct studios. The data quality is also easy to understand. Large models and intelligent driving fields need high-quality data. If the data quality of the input model is poor, the performance of the model will also be affected. One of the effective solutions to solve the data quality problem is to generate the original data through the pre-labeling of the model and then carry it out. Work labeling and then continue to strengthen learning and manual feedback to improve the quality of data labeling, or the team needs to be very clear about the data labeling process of downstream customers, and can formulate standard operating procedures so that data labeling employees can label according to it, thus improving the quality. However, how to understand the professional knowledge and knowledge coverage? Let's give three examples, which is a big challenge. It may be relatively easy to label large text models, but you must find a way to label Chinese, English and French. German, Russian, Arabic and other multilingual personnel, and how data labeling companies recruit and manage so many distributed personnel on a global scale will be no small problem. Considering an artificial intelligence application startup in the field of voice robots and digital people, startups usually don't have enough time, manpower and funds to establish a data labeling team internally. They need to find an outsourcing team to help label Chinese languages such as Sichuan accent, Cantonese accent, Shanghai accent and Northeast accent. It may be very difficult to find an excellent data annotation studio capable of these tasks in the market. If it is directly operated or subcontracted, it may take one or two months from receiving orders to recruiting, which will seriously affect the supply efficiency. Consider a more subdivided field. A start-up company specializing in legal model needs a lot of legal data annotation work. The legal field still has quite high professional requirements. Start-ups need to find data annotation suppliers who meet the following conditions and have at least a dozen understanding methods. 比特币今日价格行情网_okx交易所app_永续合约_比特币怎么买卖交易_虚拟币交易所平台

注册有任何问题请添加 微信：MVIP619 拉你进入群

文章转载注明 网址：https://netpsp.com/?id=63318