Aetech's Perspective
on Technology
Aetech's Perspective
on Technology
A Talk with Who Made It.
A Talk with Who Made It.
TECH BEHIND
TECH BEHIND
for wai-kor
for wai-kor
wai-kor is the most Korean waste sorting AI algorithm.
wai-kor is the most Korean waste sorting AI algorithm.
wai-kor's experience is a valuable asset in quickly understanding other overseas regions and creating data for them.
wai-kor's experience is a valuable asset in quickly understanding other overseas regions and creating data for them.
The perspective of Aetech's behind this technology is conveyed through the stories of the people who create the technology.
The perspective of Eight Tech behind this technology is conveyed through the stories of the people who create the technology.
To get to the point
How AI algorithms understands local waste
How AI algorithms understands local waste
Internalized
{data team}
Internalized
{data team}
Efforts to create
consistent
data quality (DQC)
Efforts to create
consistent
data quality (DQC)
Who's Behind Today?
Who's Behind Today?
Lee Do-kyung
CTO
Yoon Min-sun
Data Analysis Part
The most Korean AI's #1
In-house of the {Data Team}
In-house of the {Data Team}
"There is nothing more effective than real data"
"There is nothing
more effective
than actual data"
Lee Do-kyung
Lee Do-kyung
The strength of wai-kor is the internalization of the data team is. A system that can label according to the characteristics of waste. That was the first thing I did after joining Aetech.
The strength of Y-Core is the internalization of the data team. A system capable of labeling according to waste characteristics. This was the first thing I did when joining Eight Tech.
AI models continue to evolve. However, at the same time, quality data is needed. Especially, how to datafy new situations is important.
AI models continue to evolve. However, at the same time, quality data is needed. Especially, how to datafy new situations is important.
For example, if a PET bottle is placed on top of a foreign object, the recognition accuracy will be significantly reduced. It may be recognized as an object that is not a PET bottle or it may be recognized as background and not detected.
For example, if a PET bottle is placed on top of a foreign object, the recognition accuracy will be significantly reduced. It may be recognized as an object that is not a PET bottle or it may be recognized as background and not detected.
These issues should be quickly identified and data should be collected for the labeling tasks. In this process, auto-labeling technology using AI may be employed, but ultimately, it must go through a review by the human eye.
These issues should be quickly identified and data should be collected for the labeling tasks. In this process, auto-labeling technology using AI may be employed, but ultimately, it must go through a review by the human eye.
Aetech has its own data team which has excellent problem-solving skills and high freedom in AI development.
Eight Tech has its own data team which has excellent problem-solving skills and high freedom in AI development.
Data that understands local best
Data that understands local best
Yoon Min-sun
Yoon Min-sun
Usually, open data sets are used often. After all, it's the easiest to access. However, since they are often created abroad, it is difficult to use them directly in Korea.
Usually, open data sets are used often. After all, it's the easiest to access. However, since they are often created abroad, it is difficult to use them directly in Korea.
Internalization of the data team can be a burden for startupsit can be. In the end, it's a cost issueafter all.
Internalization of the data team can be a burden for startupsit can be. In the end, it's a cost issueafter all.
However, there are really many Korean-style experiences that only exist in Korea. For example, the shapes of cola bottles and cider bottles that are used only in Korea exist. In such cases, it is not easy to utilize datasets created abroad.
However, there are really many Korean-style experiences that only exist in Korea. For example, the shapes of cola bottles and cider bottles that are used only in Korea exist. In such cases, it is not easy to utilize datasets created abroad.
We have built a system to best understand these Korean environments and label the data for utilization. I believe that the internalization of the data team is definitely our strong point.
We have built a system to best understand these Korean environments and label the data for utilization. I believe that the internalization of the data team is definitely our strong point.
Point #1
Point #1
The internalization of the data team for the collection and analysis of the most Korean data → The secret of the most Korean AI waste algorithm.
The internalization of the data team for the collection and analysis of the most Korean data → The secret of the most Korean AI waste algorithm.
The Most Korean AI's #2
The Most Korean AI's #2
Efforts for consistent DQC
Consistent
Efforts for DQC
DQC: Data Quality Control
As if it were labeled by a single person
It's as if one person
has labeled it
Yoon Min-sun
Yoon Min-sun
Our most important goal is to ensure that no matter who is doing the work, it looks as if one person did the labeling and produces consistent results.
Our most important goal is to ensure that no matter who is doing the work, it looks as if one person did the labeling and produces consistent results.
After all, since multiple people are working, they may label the same object differently. In such cases, it can affect the consistent data quality (DQC), which is why consistent data labeling work is more important than anything else.
After all, since multiple people are working, they may label the same object differently. In such cases, it can affect the consistent data quality (DQC), which is why consistent data labeling work is more important than anything else.
So, during the labeling process, I consider it important to have a guide containing detailed information and criteria. Because no one can continuously stay 24 hours to provide training or coaching.
So, during the labeling process, I consider it important to have a guide containing detailed information and criteria. Because no one can continuously stay 24 hours to provide training or coaching.
Make decisions based on your own reference, but focus on clarity so that the results can be consistent, and we are creating guides.
Make decisions based on your own reference, but focus on clarity so that the results can be consistent, and we are creating guides.
Goal: Reduce onboarding period (2 months -> 1 month)
Goal: Reducing
onboarding period
From 2 months
To 1 month
Yoon Min-sun
Yoon Min-sun
We are seeing growth in our business and there are many new labelers joining us.
We are seeing growth in our business and there are many new labelers joining us.
In that case, it usually takes about 2 months to show a certain level of results. Education and work are carried out in parallel, and if good training is not provided at this time, it will take longer.
In that case, it usually takes about 2 months to show a certain level of results. Education and work are carried out in parallel, and if good training is not provided at this time, it will take longer.
So based on the experience we have accumulated we are preparing a new labeler training system.
So based on the experience we have accumulated we are preparing a new labeler training system.
Basic Classification Methods in addition to Common Mistakes aim to provide effective indirect experiences of failure.
Basic Classification Methods in addition to Common Mistakes aim to provide effective indirect experiences of failure.
It's not the failures caused by new cases, but just reducing failures due to repeated existing cases will greatly decrease the training period.
It's not the failures caused by new cases, but just reducing failures due to repeated existing cases will greatly decrease the training period That's why.
If the training system under preparation establishes itself effectively, the training period is expected to be shortened from 2 months to about 1 month.
If the training system under preparation establishes itself effectively, the training period is expected to be shortened from 2 months to about 1 month.
Point #2
Point #2
Creating an environment for consistent results as if labeled by one person, and a training system that reduces the adaptation period for newly joined members by half → The secret of consistent DQC.
Creating an environment for consistent results as if labeled by one person, and a training system that reduces the adaptation period for newly joined members by half → The secret of consistent DQC.
Based on the experience of wai-kor, towards the world.
Based on the experience of wai-kor, towards the world.
Lee Do-kyung
Lee Do-kyung
There are also waste types that possess unique characteristics in each region overseas. In such cases open data alone has limitations in understanding that region. It is necessary to have a data team that can collect and analyze data directly in that region.
There are also waste types that possess unique characteristics in each region overseas. In such cases open data alone has limitations in understanding that region. It is necessary to have a data team that can collect and analyze data directly in that region.
And the construction of a dataset that understands regional characteristics requires prior experience. Without prior experience, you will have to go through more trial and error, which can take more time and cost more.
And the construction of a dataset that understands regional characteristics requires prior experience. Without prior experience, you will have to go through more trial and error, which can take more time and cost more.
And this point is the strength of the technology that Aetech possesses. Processes, experiences, and know-how gained from creating and accumulating the most Korean-style waste dates are not limited to just Korea.
And this point is the strength of the technology that 8tech possesses. Processes, experiences, and know-how gained from creating and accumulating the most Korean-style waste dates are not limited to just Korea.
Understanding the region when going abroad is certainly an important asset to creating wai-kor. Therefore, minimizing trial and error and quickly building a dataset that understands the region can be done.
Understanding the region when going abroad is certainly an important asset to creating wai-kor. Therefore, minimizing trial and error and quickly building a dataset that understands the region can be done.
Thoughts while creating wai-kor are that Aetech will provide a strong foundation to thrive in more places around the world.
Thoughts while creating wai-kor are that Eight Tech will provide a strong foundation to thrive in more places around the world.
Point #3
Point #3
The traces of consideration from creating wai-kor -> a solid asset that can deeply understand overseas locals and quickly build data sets.
The traces of consideration from creating wai-kor -> a solid asset that can deeply understand overseas locals and quickly build data sets.
Aetech considers local waste for better selection. And the experiences gained in this process help to understand the area and create a better environment not only in Korea but also in the wider world.
Aetech considers local waste for better selection. And the experiences gained in this process help to understand the area and create a better environment not only in Korea but also in the wider world.