News[2023/11/18] The final evaluation has been calculated correctly.
[2023/11/16] The final evaluation is being calculated. Please wait for a while until the results appear. Once the evaluation is complete, you will receive an evaluation result email. A message prompting you to select a file will be displayed in the email, but please ignore the message.
[2023/10/23] The material of the briefing session (held on 10/11) has been released. It is available here.
[2023/10/16] The video of the second briefing session of the pre-training image dataset generation module contest has been released. https://youtu.be/qwfuhnzXn4Q
[2023/10/06] The second briefing session will be held online (Zoom) from 17:00 on October 11th. We are looking for questions regarding the competition here. Please register your participation here.
[2023/09/12] The leaderboard has been reset and you can now submit. The distribution data has also been updated accordingly, so please download it again.
[2023/09/08] The number of categories and images of each category generated by the generation module will be fixed at 100 (in other words, the total number of images generated will be 10,000). Please take this restriction into consideration when developing algorithms. Please note that you will not be able to submit until this restriction is reflected in the system. For those who have already submitted, the upper limit on the number of submissions will be reset.
[2023/09/08] The material of the briefing session (held on 9/1) has been released. It is available here.
[2023/09/04] The video of the briefing session of the pre-training image dataset generation module contest has been released. https://www.youtube.com/watch?v=aqUfRk3EO_g&ab_channel=SIGNATE
While early social implementation of artificial intelligence technology is expected, technology that accelerates the introduction of artificial intelligence technology is crucial. When introducing artificial intelligence technology there is an issue in that it takes a tremendous amount of time to tune establishment and studying of artificial intelligence models. In order to solve this problem, NEDO has a project called “Development of integration technology that becomes the core of the next-generation artificial intelligence and robots/ Research and development that expands the areas of application of artificial intelligence technology”: As a common fundamental technology that promotes early social implementation of artificial intelligence technology, the project develops technology that shortens the time to tune establishment and studying of the artificial intelligence models as well as an artificial intelligence acceleration modules using the technology.
Specifically, within the R&D themes of the above project, "Research and Development of Technology to Construct Cyber Physical Value Chains and Accelerate the Introduction of AI" and "Research and Development of Technology to Accelerate the Introduction of Artificial Intelligence Technology through Automatic Machine Learning," we have developed technologies such as the search for optimal hyperparameter (Hyperparameter Optimization, hereafter HPO) and the search for optimal neural network structures (Neural Architecture Search). The results have been compiled into an open source software called aiaccel, which is being released to the public in turn.
In order to further promote this project it is essential to induce technological development in a competitive environment which is joined by a wide range of participants including universities, corporations and overseas researchers. Therefore, we decided to hold an AI introduction acceleration module competition which will induce the creation of AI introduction acceleration modules that display excellent individual performance and versatility.
The first HPO Module Contest was held in FY2022, and despite the short duration of the contest, we were able to make it a fulfilling contest with the participation of many people. We would like to take this opportunity to once again thank everyone who participated.
In this second session, we will develop a module to mechanically generate pre-training datasets with mathematical formulas, algorithms, etc. to improve the performance of transfer learning. In research and development of AI, including the above research topics, it is essential to prepare good quality image datasets. However, in general, the construction of natural image datasets necessary for AI training requires a great deal of effort in image collection and labeling. In addition, the use of existing natural image datasets has challenges, such as restrictions on commercial use and the possibility that they may contain copyrighted images that cannot be used. Reference  reported that pre-training on a mechanically generated image dataset such as fractal images, combined with transfer learning, can ensure the same discrimination accuracy as when training on a natural image dataset. If such image datasets can be used for pre-training, it is expected to solve the issues related to the use of natural image datasets, and we look forward to your challenge as this is an excellent opportunity to try to develop a potentially revolutionary learning process for AI.
The competition consists of the first and final rounds. Ten qualifying teams which passed the first round can proceed to the final match.
|Period||September 1, 2023 to November 15, 2023.|
Note: End of the deadline for submission until 23:59:59 on November 14, 2023
Note: Final submission selection deadline until 23:59:59 on November 15, 2023
|November 24, 2023 to January 15, 2024 (scheduled).|
|Task||Development of modules for generating pre-training datasets||Development of modules for generating pre-training datasets|
|Requirement||Developing a module that works with SIGNATE server.||Developing a module that works with SIGNATE server.|
|Eligibility||Must be able to participate in the final round.|
By the end date, prepare and submit a report explaining the ingenious features of the created module.
|Test Environment||Implemented with a computational resource owned by the individual participants.||Implemented with a computational resource owned by the individual participants.|
|Assessment Environment||Evaluation of modules on servers operated by SIGNATE.||Evaluation of modules on servers operated by SIGNATE.|
|Assessment Method||Evaluated by recognition accuracy for a number of tasks (100 classes) that are subject to transfer learning.||Evaluated by recognition accuracy for a number of tasks (1000 classes) that are subject to transfer learning. |
|Leaderboard Update Frequency||Real time||Real time|
|Prize||Top ten teams with high accuracy: Eligibility to enter the final round.|
Quantitative evaluation: 1st: ¥1,200,000 2nd: ¥1,000,000 3rd: ¥800,000
Qualitative evaluation: ¥500,000 × 2 teams
Tasks for the First Round
Performance is evaluated in terms of recognition accuracy for a number of tasks (100 classes) that are subject to transfer learning.
*See Data tab for details.
Image of pre-training image dataset creation
※The number of categories and images of each category generated by the generation module must be 100(The total number of images is 10000).
Ten teams who passed the first round advance to the final round. The final round will be evaluated based on recognition accuracy for a number of tasks (1,000 classes) that are subject to transfer learning.