Whether it is the personalized recommendation of short videos, or the optimal route design for takeaway delivery, or the face recognition during payment, AI technology represented by algorithms has been applied in full swing in the consumer Internet industry.

However, when it comes to traditional industries, there are few application cases that can be talked about. Applying AI into industries outside of consumer Internet faces three major challenges: limited data resources, high customization cost, and a very long process for deployment, says Andrew Ng, founder of DeepLearning.AI. Among them, limited data resources lead to high customization cost and long process of deployment. How to secure vast and quality data has become the bottleneck for traditional industries to realize the AI transformation.

High-quality data is the foundation

The development of AI is to continuously optimize software and algorithms under the support of massive data to obtain higher algorithm accuracy. In cases where traditional industries cannot improve the quality and quantity of data, it is recommended to adopt a "data-centric" model and focus on obtaining better-quality, better-matched data. These data require professional teams to collect, clean, label, and accept before they can be obtained. Therefore, traditional industries need to cooperate with data companies to obtain high-quality vertical data, so as to promote the smooth application of AI technology to traditional industries.

Medical image recognition is a representative case of the application of AI technology in traditional industries. The AI algorithm is used to identify the CT images and assist doctors to diagnose. AI can make the diagnose process more efficient and smooth. At present, the accuracy rate of most image recognition systems can reach more than 90%, which greatly reduce the workload of doctors.

However, in other fields, like parts quality check, there is still a long way to go before AI can be fully implemented. Many manufacturers of air conditioners and washing machines will judge whether the machine is defective or not based on the sound of the machine running. This work usually requires specialist with years of experience carefully of listen the sounds. AI can be part of the process through voiceprint recognition or machine voice recognition, assisting the specialist to complete the tasks. Before launch of such AI assistant in traditional industry, large amount of the in-domain data need to be full collected and labeled, which requires the joint efforts of factories and data service companies. and at the data and algorithm level, it also requires cooperation between traditional industries and AI companies to obtain the optimal solution.

Algorithmic model adaptation

Obtaining a small amount of in-domain data is the premise of all model adaptation. The adaptation of the algorithm model can be called transfer learning. This technology is a model trained with out-of-domain data, and a small amount of in-domain data is migrated to in-domain scenarios, so that the AI model can adapt to the application scenarios in the domain.

For example, a speech recognition model trained with a large number of voice data recorded in a recording studio, the effect is tested on the voice recorded on the same channel and clean, and the recognition accuracy must be very high, but when the model is tested on the voice recorded in the vehicle environment, the effect will be rapid. decline. Therefore, it is necessary to fine-tune the model pre-trained with the original recording studio data through a small amount of in-vehicle data, so that the model can be adapted to the application scene of the in-vehicle environment.

Fine-tuning is only the simplest method of domain adaptation, and there are many methods such as adversarial training, teacher-student models, and many others. However, the premise of the task model adaptation algorithm is the in-domain. As high quality training datasets provider, Magic Data is led by a group of experienced data experts and powered with self-developed one-stop data annotation platform. The client-centric project management team responses to customized requests and delivers services to client’s needs under full compliance working process.

The development of AI should benefit all walks of life, not just the emerging Internet industry. It is also expected that various traditional industries will accelerate the pace of cooperation with AI companies, so as to seize the opportunity in the wave of AI and better lead the development of the industry.