As the culmination of Baidu AI's years of technology accumulation and business practice, Baidu Brain has developed into a world-leading artificial intelligence platform. On December 28, the Baidu Create AI Developer Conference "Baidu Brain Forum" was held. The forum focused on building a driving engine in the era of artificial intelligence, bringing the release of Baidu Brain Language and Technology, Speech and Vision and other technical products, as well as a new upgrade of Feipao in open source algorithm models, industry-level model libraries and enterprise-level AI application development.
Baidu Brain Language and Knowledge Technology is fully laid out, and the three major technology products are released
Wu Hua, chairman of Baidu's technical committee, said at the forum that after 11 years of development, Baidu has formed a complete language and knowledge technology layout, including knowledge graph, language understanding and generation technology and application system. Subsequently, Wu Hua brought the release of three major technical products: the world's first knowledge-enhancing super model Pengcheng- Baidu Wenxin, the world's largest Chinese cross-modal generation model ERNIE ViLG, and the first 10 billion parameter Sino-British dialogue pre-training generative model PLATO-XL, achieving a leading position in the fields of knowledge enhancement model, cross-modal text image generation, and human-machine dialogue.
In particular, the world's first large model of 100 billion knowledge enhancement, Baidu Wenxin, thanks to the strong combination of Pengcheng Laboratory's computing power system "Pengcheng Cloud Brain II" and the flying paddle deep learning platform, has solved a number of recognized technical problems in the training of super models, greatly improving the training efficiency and improving the model effect. Pengcheng-Baidu Wenxin has achieved the best results in more than 60 tasks such as machine reading comprehension, text classification, and semantic similarity calculation, and refreshed the benchmark on more than 30 small sample and zero sample tasks.
Based on Baidu's language and knowledge technology, Baidu has also opened up its language and knowledge open platform to various industries. It not only contains the open source dataset "Thousand Words" and the knowledge production platform "Explanation", but also develops the application-oriented capability engine platform and knowledge middle platform, as well as the intelligent document analysis platform, the intelligent dialogue customization platform, the intelligent creation platform, the translation open platform and the content moderation platform and other scenario customization platforms.
An important progress in Baidu voice technology, SMLTA2 was newly released
Speech and language are inherently closer, so the accuracy and interaction success rate of integrating the speech recognition model and the semantic model will be greatly improved. Jia Lei, chief architect of Baidu Voice, introduced the latest progress of Baidu multimodal voice interaction. SMLTA2, a streaming truncation Confomer modeling technology based on historical information abstraction proposed by Baidu, solves the computational explosion problem and storage explosion problem of traditional autocorrelation technology when recognizing long sentences, and also solves the problem of focus loss of attention model.
SMLTA2 introduces feedback through the attention feature selection mechanism of each layer of Decoder to Encoder, so that the outermost recognition result information can directly act on the coding process of each layer inside the encoder, fully extract the effective feature information through historical information abstraction, and significantly improve the various problems faced by the Transformer model from the NLP field to the speech recognition field. SMLTA2's new end-to-end modeling approach is a structural innovation in the end-to-end modeling of traditional Encoder-Decoder structures.
Finally, Jia Lei also introduced the actual commercial landing of SMLTAs. SPDB has launched a voice interaction system in a number of business halls across the country, among which the voice interaction recognition rate of bank outlets located in the Bund in Shanghai has reached 93.51%, and voice interaction has changed from completely unavailable to basically available.
Intelligent video authoring, the latest practice in computer vision
Ding Errui, director of Baidu's visual technology department, focused on the latest progress of computer vision in intelligent video creation. At this stage, video content production is shifting from UGC (user-generated content) to AIGC (AI production content).
Intelligent video creation is a multi-technology cross-integration field, for a creator, while mastering visual generation, multi-modal, 3D graphics is not realistic, but Baidu intelligent video production technology takes into account content creativity and video function creation, not only to achieve the face, human body fine processing and environmental reshaping, in the creation method, improve the stock of video and obtain new video, to ensure the adequate display and distribution of video.
Ding Errui said that the field of intelligent video creation is currently showing a vigorous development trend, and the development of technology has brought about changes in production tools, which, once combined with other production factors, will bring endless imagination.
The flying propeller industry-level platform has been upgraded to make the threshold for AI applications lower
In addition to the integration and innovation of technology, in terms of tools and platforms, Feipao has upgraded from the aspects of open source algorithm models, industry-level model libraries and enterprise-level AI application development, and continues to reduce the threshold of AI applications.
Bi Ran, an outstanding architect at Baidu, shared that there are currently more than 400 industry-level open source algorithm models officially supported by Baidu Feipao, covering many deep learning application fields such as computer vision, natural language processing, speech and recommendation. This comprehensive coverage allows developers to quickly find the model they need. And these industry-grade model libraries enable full-process support for training and deployment, and development kits support flexible configuration and tuning.
Bi Ran also introduced in detail the latest industrial practice example library launched by Flying Propeller. For ai applications in actual industrial scenarios, the sample library provides a complete code implementation, covering the whole process of industrial landing such as task analysis, algorithm selection, model training and optimization, inference deployment and result visualization, so that developers can quickly get started and use what they have learned.
Xin Zhou, Director of Baidu Intelligent Cloud AI Product R&D Department, introduced the effective help of Flying Propeller Enterprise Edition AI Development Dual Platform EasyDL and BML in helping developers quickly improve AI development efficiency and resource use efficiency.
At present, Flying Propeller Enterprise Edition has become the most widely used and landed AI development platform. Based on the dual-platform development model, it meets the needs of AI application developers and AI algorithm developers. Based on the Flying Propeller Inference Deployment Toolchain, Flying Propeller Enterprise Edition has tested and verified the combination of 9345 model chips for developers, which can cover 95% of the adaptation requirements and save 97% of the developer's self-adaptation development time. PaddleSlim combined with a fully automated model combination compression algorithm can improve the performance of inference by 3 to 5 times when the accuracy loss is controlled within 1%, and the introduction of the intelligent edge console greatly improves the efficiency of module and system integration, and the integration cycle can be shortened from days to five minutes.
Zhu Yong, senior director of Baidu's Knowledge Graph Department and Big Data Department, introduced in detail how the industrial data intelligence engine built by Baidu can reduce the threshold for AI application in the industrial field.
Zhu Yong said that with the deepening of the process of industrial digitalization, the application trend of big data has gradually developed from business data to data intelligence, and industrial data intelligence has broad prospects, and opportunities and challenges coexist. Based on the leading artificial intelligence big data technology, Baidu has created a complete set of industrial data intelligence engines for industrial scenarios. It docks down to the big data platform to achieve data governance, supports the needs of various types of business scenarios upwards, and empowers different industries such as power energy, steel, chemicals, and automobile manufacturing. At the heart of Baidu's Industrial Data Intelligence Engine is a series of reusable industrial models and core components that support customized model building, including data analysis, data processing, AI algorithms, and industrial mechanisms.
Taking the steel industry as an example, in order to ensure that the galvanized sheet has good mechanical properties, it is necessary to make necessary adjustments to the galvanizing process parameters according to the raw material information of the steel plate. Based on Baidu industrial data intelligence engine, mechanical performance prediction can be carried out, reaching an accuracy rate of more than 90%, and the product qualification rate reaches 99% through the optimization of process parameters. By applying this solution, enterprise customers can not only improve product quality, but also significantly reduce commissioning time compared to traditional manual experience-based methods, thereby improving production efficiency.
Finally, Zhu Yong stressed that in the context of the accelerated penetration of big data and AI into industry, the prospect of industrial data intelligence can be expected. Baidu looks forward to working with more developers to jointly help industrial intelligent upgrading.
Next, Baidu Brain will continue to evolve, continue to promote the improvement of AI technology, through the integration of innovation, so that AI capabilities are getting stronger and stronger, at the same time, reduce the threshold of AI applications, make the landing of technology simpler, help more and more industries and enterprises to improve efficiency with AI, create value.