The 1st Koh Young AI Competition

Date: September 13 ~ October 19, 2023
HTTP://KohYoungAi.iccas.org
Organized by Institute of Control Robotics and Systems (ICROS)
Sponsored by Koh Young Technology

Purpose

Recently, artificial intelligence (AI) has been widely applied to improve operational efficiency in various industrial fields. In line with this, ICROS and Koh Young (https://kohyoung.com/kr), the world-leading SMT inspection solution provider, jointly host AI Challenge to encourage students to compete for essential decision problems in the SMT manufacturing process.

Participants

All graduate and undergraduate students with AI development experience can participate in the competition. The maximum team size is 5.
The competition problem will be about practical and effective learning of image data collected from industrial sites.

Registration

1. Download and fill out the registration form: Click & Download the form
2. Send the form to conference@icros.org and sooheehan@postech.ac.kr.

Main schedule

Registration opening: Aug. 8 – Sep. 11, 2023
Competition: Sep. 13 – ~~Oct. 4, 2023~~ Oct. 10, 2023
Award ceremony: October 19, 2023

All winners must register for ICCAS 2023. They also have opportunities to present their algorithms in the special session and are required to open their sources. The oral presentations of the 1st, 2nd and 3rd place winners will be recorded and uploaded to YouTube.

Competition schedule is subject to slight change due to various circumstances.

Distributed data

All training data will be distributed in the form of a single compressed file. Participants will receive this data file through personal e-mail addresses. The file name is given as Koh_Young_AI_data.tar. This file consists of 1000 folders named from 1 to 1000. Each folder has 150 images.

Problem: Large Scale Optical Text Incremental Learning

All images of each folder are assigned one designated label. It means that each folder is associated with a certain label. For convenience, it can be said that all images of the folder named 59 are labeled with 59. Design a total of ten network models that take one image of a prescribed pixel size as input and return its label as output. A first network model is trained on data labeled 1 to 100, and tested on data from the same classes. A second network model initialized with the first one is additionally trained on data labeled 101 to 200 and tested on data labeled 1 to 200. In the same way, a third network model initialized with the second one is additionally trained on data labeled 201 to 300 and tested on data labeled 1 to 300. Repeating this produces a total of 10 network models and the 10 resulting test scores.

Evaluation metric: Announced to all teams.

Final submission

Participants are required to submit their codes that meet regulations.
Template codes will be provided when the competition starts.

Prizes (8 teams or more)

- 1st place: 5,000,000 KRW
- 2nd place: 3,000,000 KRW
- 3rd place: 1,000,000 KRW
- 4th place or lower: 500,000 KRW for each

Announcement (last updated: Oct. 19)

– Sep. 15 : New template files are provided, which are marked for distinguishing what you can change from what you cannot.
You must submit a code file (.ipynb) and a model file (.pt or .h5).
The model file refers to the one that has completed all incremental learning.

– Sep. 18 : After loading all training image data, you can perform data-augmentation based on them.
The preprocessing code for data augmentation should go where you can change codes.
Pretrained models are not allowed.

– Sep. 18 : In this competition, your task is to classify 100 classes at a time and perform this 10 times.
For each learning process, you should not use the previous data.
For example, when you learn classes from 201 to 300, you cannot use image data
corresponding to ones from 1 to 200, in any way. This competition aims to efficiently suppress
catastrophic forgetting.

– Sep. 19 : There is a bug in the template code. Please modify the code as follows:
end=100*div_idx+100 -> end=100*div_idx+101
To learn the next 100 classes, you should retrain the previously trained network
with the corresponding image data, while avoiding catastrophic forgetting.
You can modify code in the last block. But, data loading and training must
be conducted in accordance with regulations.

– Oct. 3 : At the request of many participants, the competition deadline has been extended to October 10th.
As previously announced, all participants will be given two opportunities to submit their resulting files.
For fairness, files submitted already will be canceled.
Those who have already submitted their results, also have two opportunities.

– Oct. 4 : The participants are required to run the validation function for self-assessment.

– Oct. 16 : Thank you for your active participation in the competition.
Finally, five teams are selected based on performance measured through subsequent testing.

Gold/Silver Medals: MVCL and KNU AIMotion
Bronze Medals: RML, Air lab, and SPAC

Gold and silver medals will be awarded based on the presentation evaluations this coming Thursday.

We hope to see you next year with even more improved skills.
Thank you.

– Oct. 19 : The award ceremony was held at the banquet of ICCAS 2023.
The final winning teams are as follows:

1st prize winning teams : MVCL and KNU AIMotion
3rd prize winning teams : RML, Air lab, and SPAC

Thank all participants for their passion and effort.
See you next ICCAS.

Contact: If you have any questions, please send your inquiry to Prof. Soohee Han at sooheehan@postech.ac.kr by e-mail.