The Social Computing Innovation Competition is jointly organized by the China Artificial Intelligence Association (CAAI) and the Big Data Development and Management Bureau of Zhejiang Province. Co-hosted by the CAAI Committee on Social Computing and Social Intelligence and the School of Public Administration at Zhejiang University, the competition aims to promote digital government development, leverage the foundational role and innovative potential of data, and support the province's overall digital reform.

The construction of a 'brain,' serving as a breakthrough in advancing the depth of digital reform, is the focal point. The competition will concentrate on the development of 'brain' algorithms, models, and various cross-scenario application designs. Emphasizing data empowerment and reinforcing practical effectiveness, it aims to provide robust support for Zhejiang Province's digital reform. Additionally, the competition hopes that its outcomes can drive the modernization of the national governance system and capabilities, thereby offering strong support for the construction of a digital China.

Preliminary Topic 1: Bipolar Obstacle Detection

• Task Description

Bipolar disorder is a type of mood disorder, also known as Bipolar Disorder (BD) or Bipolar Affective Disorder, characterized by episodes of mania and depression. The etiology is currently unclear, with biological, psychological, and social environmental factors all playing a part in its development. Recent research has found that genetic factors, environmental or stress factors, their interactions, and the timing of these interactions all have significant impacts on the development of bipolar disorder. Clinically, it can be categorized into depressive, manic, or mixed episodes. Bipolar disorder detection involves predicting whether a patient has bipolar disorder or whether the treatment for bipolar disorder is effective, using medical diagnostic data, which includes medical imaging and gut data. Due to the lack of medical samples and the abundance of features, selecting appropriate features to integrate bimodal features and training a suitable classifier for model prediction is of great practical need and medical significance. Participating teams need to build a classification model based on the provided dataset, compile it using the SOPHON MLIR open-source compiler, and submit the compiled bmodel file, the original model file, and the inference script code.

• Task Objective

The task requires participating teams to complete supervised learning with few samples and many features.

Preliminary Topic 2: Social Relationship Extraction

• Task Description

Relation extraction is an important subject in social computing and a fundamental task in information extraction, as well as a crucial step in studying the laws of social operation. Relation extraction involves extracting triplets (subject, relation, object) from a text. The main tasks are to identify the subject and object in the text (entity recognition task) and to determine the relationship between these two entities (relationship classification). Moreover, relation extraction can support the automatic construction of knowledge graphs, search engines, question answering, and other downstream tasks. Participating teams are required to use a designated pre-trained relation extraction model and dataset, compile it using the SOPHON MLIR open-source compiler, and submit the compiled bmodel file, the original model file, and the inference script code.

• Task Objective

Participating teams need to select the most suitable model from five relation extraction models, compile it into a bmodel using the MLIR compiler, and add pre-processing and post-processing programs to implement the application of the model, which should be able to correctly process the data.

Note: The preliminary competition dataset and login portal for this track will be available in mid-March.

Semi-Final Topic 1: Bipolar Obstacle Detection

• Task Description

For the bipolar disorder detection task from Preliminary Topic 1, participating teams need to compile on the SOPHON.NET using the SOPHON MLIR open-source compiler, and submit the compiled bmodel file, the original model file, and the inference script code.

• Task Objective

The task requires participating teams to complete supervised learning with few samples and many features.

Semi-Final Topic 2: Social Relationship Extraction Model Deployment

• Task Description

For the relation extraction task from Preliminary Topic 2, compile using the SOPHON MLIR open-source compiler, perform quantization and tuning, and deploy the model on the SOPHON.NET TPU to test its inference performance.

• Task Objective

Participating teams need to select the most suitable model from five relation extraction models, compile it into a bmodel using the MLIR compiler, and add pre-processing and post-processing programs to implement the application of the model, which should be able to correctly process the data.

Final Topic: Face Recognition Model Deployment Based on TPU-MLIR

• Task Description

Face recognition is a classic subject in social relationship identification and a preliminary step in studying human social attributes. Face recognition is a biometric technology that identifies individuals based on facial feature information. It involves the collection of images or video streams containing faces using cameras or video cameras, the automatic detection and tracking of faces in images, and a series of related technologies for facial recognition of detected faces, commonly referred to as facial recognition or face recognition.

The TPU-MLIR project is a TPU compiler engineering for SOPHON Intelligent processors. This project provides a complete toolchain that can convert pre-trained neural networks from different frameworks into binary files that can be efficiently computed on SOPHON TPU.

• Task Objective

Participants can use the provided pre-trained models and datasets to train their face recognition models, compile, quantify, and tune using the SOPHON MLIR open-source compiler.

• Registration & Team Formation (2023.04.15)

 

The registration portal is now open. Contestants can log in to the Mo (momodel.cn) platform to register for the competition.

• Preliminary Stage (2023.04.17-2023.05.26)

 

During the preliminary stage, to fully explore the advantages of interdisciplinary integration and leverage the strengths of researchers from different fields, the competition topics include bidirectional obstacle detection and social relationship extraction. Contestants can choose either topic. Contestants must log in to the Mo platform to write their code and submit their results. On 2023.05.30, the results of the preliminary round will be announced and displayed on the Mo platform.

• Semi-Final Stage (2023.06.01-2023.06.20)

 

Participants need to complete model conversion and inference tasks on the SOPHON.NET platform. In accordance with the rules, a certain number of outstanding teams from the two competition tracks will be selected to enter the finals. On 2023.06.25, the results of the semi-finals will be announced and displayed on both the SOPHON Cloud platform and the Mo platform.

• Final Stage (2023.07.07-2023.07.09)

 

Teams that make it to the finals will be invited to the conference site to participate in the final competition. The final stage will provide new competition topics and will be conducted in the form of a Hackathon (lasting a total of 36 hours). The organizing committee of the competition will judge and award several winning teams on-site, and present them with honorary certificates and prize money.

The only official registration platform for this competition is Mo platform (https://momodel.cn/competition). This page is for display and introduction purposes only. Participants interested in this competition are kindly requested to proceed to the Mo platform for registration.

Judging Rules

1. The total score for the competition is out of 100 points, assessed automatically by the system.
2. After all automatic system assessments, two designated judges appointed by the chief referee will review and confirm all scores. Any modifications or issues discovered during this review must be explained to the chief referee and duly recorded.
3. Following the confirmation, a comprehensive score sheet for participating teams will be generated and signed by the chief referee. This includes the work task assignment, on-site records, confirmation sheets, and other relevant physical documents, which will be handed over to the organizing committee.
4. The final ranking of participating teams will be based on the competition scores in descending order, unrelated to the time taken to complete the tasks.
5. Once the chief referee formally submits the competition work assessment results and verifies their accuracy, the final scores will be announced during the platform's closing ceremony.

Team Formation Rules

1. Participating teams can only be formed by team leaders who have already registered. Registered team members need to join the team formed by their respective team leader to participate.
2. Registered team leaders can manage their teams by logging into the competition website. This includes creating teams, inviting members, managing team members, and disbanding teams.
3. Each participating team can have a maximum of 5 members.
4. One registrant can only belong to one participating team.
5. After the registration period ends, no further team creation or changes to team members (such as inviting, removing, joining, leaving, or disbanding teams) will be allowed.