Since the beginning of the year, the artificial intelligence large language model (LLM) has set off a new round of global technology competition, and global technology giants have started the "Hundred Model War". When the big language model is profoundly changing human production and lifestyle, how to further release its potential has become a concern of the industry and the starting point of the proposition of the 2023 language and intelligent technology competition.
On May 17, the 2023 Language and Intelligence Technology Competition was officially launched. The competition is jointly sponsored by the China Computer Federation (CCF) and the Chinese Information Society of China (CIPS). Organized by the Society Evaluation Working Committee. This year's competition set up two tracks, "Large Language Model Ability Evaluation" and "Video Semantic Understanding", based on practical problems, discovering innovative talents, and boosting technological development.
Keeping up with the technology trend proposition, the global talent gathering "reveals the list"
The Language and Intelligence Technology Competition has been held for 5 consecutive sessions since 2018. With task design for real application scenarios and data sets derived from real scenarios, it has attracted attention from all walks of life in industry, academia and research, and has become the most authoritative and popular Chinese natural language in the world. Handle one of the events.
Previous competitions have successively organized evaluation tasks such as reading comprehension, man-machine dialogue, semantic analysis, and information extraction, covering important frontier topics in the fields of natural language processing and artificial intelligence. The development of intelligent applications is of great significance. Each contest topic is like a hero post, attracting more than 2,000 teams to "unveil the rankings", 80% of which come from many top universities and technology companies around the world, covering finance, Internet, media, communications, construction machinery, energy, biology, etc. industry.
At present, the big language model is a product of "big data + big computing power + strong algorithm". After pre-training based on trillion-level data sets, it can meet diverse needs. It is regarded as a milestone technology for AI to move towards AGI (artificial general intelligence) . Keeping up with technological trends, this year's competition set up two tracks, "Large Language Model Ability Evaluation" and "Video Semantic Understanding", aiming to cooperate with global innovative talents to contribute to the development and application of language and intelligence technologies.
Looking for "National Evaluation Officer" to build a large-scale model ability evaluation system
Unlike previous competitions, which were mainly for professional AI developers, the first track of this competition, "Large Language Model Ability Evaluation", invites all users . The track expects contestants to formulate evaluation plans and data examples from the dimensions of underlying abilities (generation, logic, etc.), special abilities (creation, question-and-answer, etc.), and application abilities in real scenarios, and build an ability evaluation system for large language models ( See examples below).
Evaluation system example
The setting of this competition question is particularly exciting for individual users and small and medium-sized development teams. On the one hand, the advent of the large language model Zhatui has caused problems such as homogeneity, and it is urgent to propose a comprehensive and effective evaluation method; on the other hand, the number of model parameters has surged to trillions, and the cost of single training is high. Businesses can afford it. By participating in the language and intelligence technology competition, contestants only need to start from their own understanding of the large model and establish logical and smooth evaluation dimensions and evaluation standards to participate in technological change at low cost. As the organizer, Baidu will provide all contestants with an invitation code for a new generation of knowledge-enhanced large language model Wenxinyiyan to help contestants better establish a large-scale model evaluation system.
This also means that the first track has almost "zero threshold". There is no age limit, no professional limit, and no code foundation. As long as there is reason and evidence, you can serve as a "national evaluation officer" to help people understand the adaptability and limitations of large language models in different scenarios, so as to make them more secure and controllable.
Explore the "ceiling" of multimodal capabilities and strengthen video semantic understanding
The second track "Video Semantic Understanding" focuses on professional AI developers . The evaluation task takes Internet video as input, and needs to be based on perceptual content analysis (such as face recognition, OCR recognition, speech recognition, etc.) Knowledge, NLP, voice and other multi-modal information, combined with knowledge map calculation and reasoning, generates semantic tags of multi-knowledge dimensions for videos.
This task is a preliminary exploration of AI's multi-modal capabilities and AGI. Multimodality is considered by the industry to be the next development direction of large language models. Similar to how humans obtain most of the information based on the visual system, AI is also moving from single-modal intelligence such as text, voice, and vision to AGI that integrates multiple modalities. In the direction of development, GPT-4's image recognition ability and Wenxinyiyan's literary image ability are both manifestations of multimodal capabilities.
Competitors participating in the second track will get the baseline system based on the flying paddle platform provided by Baidu , so that they can get started quickly and achieve the best competition conditions. At the same time, players can also use the online programming environment based on AI Studio, an artificial intelligence learning and training community of Baidu Fei Paddle , to obtain free GPU computing power support, break the shackles of computing power, and continuously deepen their understanding of AI multimodal capabilities.
The "Hundred Model War" is now in full swing, and AGI is no longer far away. As Baidu CTO Wang Haifeng said, "The versatility of large models is becoming stronger and stronger, and AGI has been realized to a certain extent, but the direction of our efforts is to bring value to human beings through AI." The two tracks of this competition not only unite the most extensive participants in the AI era to build a comprehensive and scientific evaluation system, but also encourage professional AI developers to firmly move towards the next technological high point, so that AI can serve human beings better life and social development. At the same time, Baidu also continues to promote the "5 million AI talents in 5 years" plan through competitions, school-enterprise cooperation, etc., and continues to contribute to the construction of national strategic scientific and technological capabilities.
From now on, the registration channel for the 2023 Language and Intelligence Technology Competition is open. For details, please visit the official website of the competition. The competition has also prepared a generous prize pool, and the winning teams will have the opportunity to present their works with experts in many fields at the 2023 Language and Intelligence Summit Forum.
- Official website link